Add/update custom_nodes

custom_nodes/comfyui-segment-anything-2/.gitattributes

@@ -0,0 +1,2 @@
+# Auto detect text files and perform LF normalization
+* text=auto

custom_nodes/comfyui-segment-anything-2/.github/workflows/publish.yml

@@ -0,0 +1,25 @@
+name: Publish to Comfy registry
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - main
+    paths:
+      - "pyproject.toml"
+
+permissions:
+  issues: write
+
+jobs:
+  publish-node:
+    name: Publish Custom Node to registry
+    runs-on: ubuntu-latest
+    if: ${{ github.repository_owner == 'kijai' }}
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v4
+      - name: Publish Custom Node
+        uses: Comfy-Org/publish-node-action@v1
+        with:
+          ## Add your own personal access token to your Github Repository secrets and reference it here.
+          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}

custom_nodes/comfyui-segment-anything-2/.gitignore

@@ -0,0 +1,9 @@
+.DS_Store
+*pyc
+.vscode
+__pycache__
+*.egg-info
+*.bak
+checkpoints
+results
+backup

custom_nodes/comfyui-segment-anything-2/.tracking

@@ -0,0 +1,43 @@
+.gitattributes
+.github/workflows/publish.yml
+.gitignore
+LICENSE
+__init__.py
+example_workflows/florence_segment_2.json
+example_workflows/image_batch_bbox_segment.json
+example_workflows/points_segment_video_example.json
+load_model.py
+nodes.py
+pyproject.toml
+readme.md
+sam2/__init__.py
+sam2/automatic_mask_generator.py
+sam2/modeling/__init__.py
+sam2/modeling/backbones/__init__.py
+sam2/modeling/backbones/hieradet.py
+sam2/modeling/backbones/image_encoder.py
+sam2/modeling/backbones/utils.py
+sam2/modeling/memory_attention.py
+sam2/modeling/memory_encoder.py
+sam2/modeling/position_encoding.py
+sam2/modeling/sam/__init__.py
+sam2/modeling/sam/mask_decoder.py
+sam2/modeling/sam/prompt_encoder.py
+sam2/modeling/sam/transformer.py
+sam2/modeling/sam2_base.py
+sam2/modeling/sam2_utils.py
+sam2/sam2_image_predictor.py
+sam2/sam2_video_predictor.py
+sam2/utils/__init__.py
+sam2/utils/amg.py
+sam2/utils/misc.py
+sam2/utils/transforms.py
+sam2_configs/__init__.py
+sam2_configs/sam2.1_hiera_b+.yaml
+sam2_configs/sam2.1_hiera_l.yaml
+sam2_configs/sam2.1_hiera_s.yaml
+sam2_configs/sam2.1_hiera_t.yaml
+sam2_configs/sam2_hiera_b+.yaml
+sam2_configs/sam2_hiera_l.yaml
+sam2_configs/sam2_hiera_s.yaml
+sam2_configs/sam2_hiera_t.yaml

custom_nodes/comfyui-segment-anything-2/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

custom_nodes/comfyui-segment-anything-2/__init__.py

@@ -0,0 +1,3 @@
+from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
+
+__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]

custom_nodes/comfyui-segment-anything-2/example_workflows/florence_segment_2.json

@@ -0,0 +1,579 @@
+{
+  "last_node_id": 102,
+  "last_link_id": 239,
+  "nodes": [
+    {
+      "id": 83,
+      "type": "LoadImage",
+      "pos": [
+        -6,
+        40
+      ],
+      "size": {
+        "0": 315,
+        "1": 314
+      },
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            196
+          ],
+          "shape": 3,
+          "slot_index": 0
+        },
+        {
+          "name": "MASK",
+          "type": "MASK",
+          "links": null,
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LoadImage"
+      },
+      "widgets_values": [
+        "truck.jpg",
+        "image"
+      ]
+    },
+    {
+      "id": 66,
+      "type": "DownloadAndLoadSAM2Model",
+      "pos": [
+        -34,
+        -171
+      ],
+      "size": {
+        "0": 351.7801513671875,
+        "1": 130
+      },
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "outputs": [
+        {
+          "name": "sam2_model",
+          "type": "SAM2MODEL",
+          "links": [
+            236
+          ],
+          "shape": 3,
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DownloadAndLoadSAM2Model"
+      },
+      "widgets_values": [
+        "sam2_hiera_small.safetensors",
+        "single_image",
+        "cuda",
+        "bf16"
+      ]
+    },
+    {
+      "id": 84,
+      "type": "ImageAndMaskPreview",
+      "pos": [
+        958,
+        -293
+      ],
+      "size": {
+        "0": 667.9199829101562,
+        "1": 541.2733154296875
+      },
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 192
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 238,
+          "slot_index": 1
+        }
+      ],
+      "outputs": [
+        {
+          "name": "composite",
+          "type": "IMAGE",
+          "links": null,
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ImageAndMaskPreview"
+      },
+      "widgets_values": [
+        1,
+        "255, 0, 0",
+        false
+      ]
+    },
+    {
+      "id": 72,
+      "type": "ImageResizeKJ",
+      "pos": [
+        353,
+        127
+      ],
+      "size": {
+        "0": 315,
+        "1": 242
+      },
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 196
+        },
+        {
+          "name": "get_image_size",
+          "type": "IMAGE",
+          "link": null
+        },
+        {
+          "name": "width_input",
+          "type": "INT",
+          "link": null,
+          "widget": {
+            "name": "width_input"
+          }
+        },
+        {
+          "name": "height_input",
+          "type": "INT",
+          "link": null,
+          "widget": {
+            "name": "height_input"
+          }
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            192,
+            210,
+            226,
+            237
+          ],
+          "shape": 3,
+          "slot_index": 0
+        },
+        {
+          "name": "width",
+          "type": "INT",
+          "links": null,
+          "shape": 3
+        },
+        {
+          "name": "height",
+          "type": "INT",
+          "links": null,
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ImageResizeKJ"
+      },
+      "widgets_values": [
+        768,
+        512,
+        "nearest-exact",
+        false,
+        2,
+        0,
+        0
+      ]
+    },
+    {
+      "id": 99,
+      "type": "PreviewImage",
+      "pos": [
+        1044,
+        -744
+      ],
+      "size": {
+        "0": 530.9268798828125,
+        "1": 363.34893798828125
+      },
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 226
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "PreviewImage"
+      }
+    },
+    {
+      "id": 90,
+      "type": "PreviewImage",
+      "pos": [
+        422,
+        -800
+      ],
+      "size": {
+        "0": 568.406494140625,
+        "1": 384.9489440917969
+      },
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 200
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "PreviewImage"
+      }
+    },
+    {
+      "id": 93,
+      "type": "Florence2toCoordinates",
+      "pos": [
+        399,
+        -314
+      ],
+      "size": {
+        "0": 210,
+        "1": 78
+      },
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "data",
+          "type": "JSON",
+          "link": 204
+        }
+      ],
+      "outputs": [
+        {
+          "name": "coordinates",
+          "type": "STRING",
+          "links": [],
+          "shape": 3,
+          "slot_index": 0
+        },
+        {
+          "name": "bboxes",
+          "type": "BBOX",
+          "links": [
+            239
+          ],
+          "shape": 3,
+          "slot_index": 1
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "Florence2toCoordinates"
+      },
+      "widgets_values": [
+        ""
+      ]
+    },
+    {
+      "id": 87,
+      "type": "Florence2Run",
+      "pos": [
+        -85,
+        -796
+      ],
+      "size": {
+        "0": 400,
+        "1": 304
+      },
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 210,
+          "slot_index": 0
+        },
+        {
+          "name": "florence2_model",
+          "type": "FL2MODEL",
+          "link": 197,
+          "slot_index": 1
+        }
+      ],
+      "outputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "links": [
+            200
+          ],
+          "shape": 3,
+          "slot_index": 0
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "links": null,
+          "shape": 3
+        },
+        {
+          "name": "caption",
+          "type": "STRING",
+          "links": null,
+          "shape": 3,
+          "slot_index": 2
+        },
+        {
+          "name": "data",
+          "type": "JSON",
+          "links": [
+            204
+          ],
+          "shape": 3,
+          "slot_index": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "Florence2Run"
+      },
+      "widgets_values": [
+        "wheel",
+        "caption_to_phrase_grounding",
+        true,
+        false,
+        1024,
+        3,
+        true,
+        ""
+      ]
+    },
+    {
+      "id": 102,
+      "type": "Sam2Segmentation",
+      "pos": [
+        440,
+        -120
+      ],
+      "size": [
+        314.5386123916544,
+        162
+      ],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "sam2_model",
+          "type": "SAM2MODEL",
+          "link": 236
+        },
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 237
+        },
+        {
+          "name": "bboxes",
+          "type": "BBOX",
+          "link": 239
+        },
+        {
+          "name": "coordinates_positive",
+          "type": "STRING",
+          "link": null,
+          "widget": {
+            "name": "coordinates_positive"
+          }
+        },
+        {
+          "name": "coordinates_negative",
+          "type": "STRING",
+          "link": null,
+          "widget": {
+            "name": "coordinates_negative"
+          }
+        }
+      ],
+      "outputs": [
+        {
+          "name": "mask",
+          "type": "MASK",
+          "links": [
+            238
+          ],
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "Sam2Segmentation"
+      },
+      "widgets_values": [
+        true,
+        "",
+        "",
+        true
+      ]
+    },
+    {
+      "id": 88,
+      "type": "DownloadAndLoadFlorence2Model",
+      "pos": [
+        -470,
+        -777
+      ],
+      "size": {
+        "0": 315,
+        "1": 106
+      },
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "outputs": [
+        {
+          "name": "florence2_model",
+          "type": "FL2MODEL",
+          "links": [
+            197
+          ],
+          "shape": 3,
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DownloadAndLoadFlorence2Model"
+      },
+      "widgets_values": [
+        "microsoft/Florence-2-base",
+        "fp16",
+        "sdpa"
+      ]
+    }
+  ],
+  "links": [
+    [
+      192,
+      72,
+      0,
+      84,
+      0,
+      "IMAGE"
+    ],
+    [
+      196,
+      83,
+      0,
+      72,
+      0,
+      "IMAGE"
+    ],
+    [
+      197,
+      88,
+      0,
+      87,
+      1,
+      "FL2MODEL"
+    ],
+    [
+      200,
+      87,
+      0,
+      90,
+      0,
+      "IMAGE"
+    ],
+    [
+      204,
+      87,
+      3,
+      93,
+      0,
+      "JSON"
+    ],
+    [
+      210,
+      72,
+      0,
+      87,
+      0,
+      "IMAGE"
+    ],
+    [
+      226,
+      72,
+      0,
+      99,
+      0,
+      "IMAGE"
+    ],
+    [
+      236,
+      66,
+      0,
+      102,
+      0,
+      "SAM2MODEL"
+    ],
+    [
+      237,
+      72,
+      0,
+      102,
+      1,
+      "IMAGE"
+    ],
+    [
+      238,
+      102,
+      0,
+      84,
+      1,
+      "MASK"
+    ],
+    [
+      239,
+      93,
+      1,
+      102,
+      2,
+      "BBOX"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 0.7627768444385467,
+      "offset": [
+        564.3268832902941,
+        896.4031145502903
+      ]
+    }
+  },
+  "version": 0.4
+}

custom_nodes/comfyui-segment-anything-2/example_workflows/image_batch_bbox_segment.json

@@ -0,0 +1,766 @@
+{
+  "last_node_id": 30,
+  "last_link_id": 58,
+  "nodes": [
+    {
+      "id": 2,
+      "type": "DownloadAndLoadSAM2Model",
+      "pos": [
+        119,
+        52
+      ],
+      "size": {
+        "0": 315,
+        "1": 130
+      },
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "outputs": [
+        {
+          "name": "sam2_model",
+          "type": "SAM2MODEL",
+          "links": [
+            9
+          ],
+          "shape": 3,
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DownloadAndLoadSAM2Model"
+      },
+      "widgets_values": [
+        "sam2_hiera_base_plus.safetensors",
+        "single_image",
+        "cuda",
+        "bf16"
+      ]
+    },
+    {
+      "id": 13,
+      "type": "DownloadAndLoadFlorence2Model",
+      "pos": [
+        105,
+        -299
+      ],
+      "size": {
+        "0": 315,
+        "1": 106
+      },
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "lora",
+          "type": "PEFTLORA",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "florence2_model",
+          "type": "FL2MODEL",
+          "links": [
+            23
+          ],
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DownloadAndLoadFlorence2Model"
+      },
+      "widgets_values": [
+        "microsoft/Florence-2-large",
+        "fp16",
+        "sdpa"
+      ]
+    },
+    {
+      "id": 26,
+      "type": "MaskToImage",
+      "pos": [
+        1161,
+        280
+      ],
+      "size": {
+        "0": 210,
+        "1": 26
+      },
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 43
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            44
+          ],
+          "shape": 3,
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "MaskToImage"
+      }
+    },
+    {
+      "id": 25,
+      "type": "ImageCompositeMasked",
+      "pos": [
+        1124,
+        364
+      ],
+      "size": {
+        "0": 315,
+        "1": 146
+      },
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "destination",
+          "type": "IMAGE",
+          "link": 55,
+          "slot_index": 0
+        },
+        {
+          "name": "source",
+          "type": "IMAGE",
+          "link": 44
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 45
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            56
+          ],
+          "shape": 3,
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ImageCompositeMasked"
+      },
+      "widgets_values": [
+        0,
+        0,
+        false
+      ]
+    },
+    {
+      "id": 29,
+      "type": "VHS_VideoCombine",
+      "pos": [
+        1486,
+        58
+      ],
+      "size": [
+        772.7946166992188,
+        859.3206163194444
+      ],
+      "flags": {},
+      "order": 10,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 56,
+          "slot_index": 0
+        },
+        {
+          "name": "audio",
+          "type": "VHS_AUDIO",
+          "link": null
+        },
+        {
+          "name": "meta_batch",
+          "type": "VHS_BatchManager",
+          "link": null
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "Filenames",
+          "type": "VHS_FILENAMES",
+          "links": null,
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VHS_VideoCombine"
+      },
+      "widgets_values": {
+        "frame_rate": 3,
+        "loop_count": 0,
+        "filename_prefix": "AnimateDiff",
+        "format": "video/h264-mp4",
+        "pix_fmt": "yuv420p",
+        "crf": 19,
+        "save_metadata": true,
+        "pingpong": false,
+        "save_output": false,
+        "videopreview": {
+          "hidden": false,
+          "paused": false,
+          "params": {
+            "filename": "AnimateDiff_00002.mp4",
+            "subfolder": "",
+            "type": "temp",
+            "format": "video/h264-mp4",
+            "frame_rate": 3
+          }
+        }
+      }
+    },
+    {
+      "id": 11,
+      "type": "VHS_LoadVideo",
+      "pos": [
+        76,
+        274
+      ],
+      "size": [
+        235.1999969482422,
+        429.0311089409722
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "meta_batch",
+          "type": "VHS_BatchManager",
+          "link": null
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            28,
+            37
+          ],
+          "shape": 3,
+          "slot_index": 0
+        },
+        {
+          "name": "frame_count",
+          "type": "INT",
+          "links": null,
+          "shape": 3
+        },
+        {
+          "name": "audio",
+          "type": "VHS_AUDIO",
+          "links": null,
+          "shape": 3
+        },
+        {
+          "name": "video_info",
+          "type": "VHS_VIDEOINFO",
+          "links": null,
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VHS_LoadVideo"
+      },
+      "widgets_values": {
+        "video": "katana_02.mp4",
+        "force_rate": 0,
+        "force_size": "Disabled",
+        "custom_width": 512,
+        "custom_height": 512,
+        "frame_load_cap": 16,
+        "skip_first_frames": 0,
+        "select_every_nth": 5,
+        "choose video to upload": "image",
+        "videopreview": {
+          "hidden": false,
+          "paused": false,
+          "params": {
+            "frame_load_cap": 16,
+            "skip_first_frames": 0,
+            "force_rate": 0,
+            "filename": "katana_02.mp4",
+            "type": "input",
+            "format": "video/mp4",
+            "select_every_nth": 5
+          }
+        }
+      }
+    },
+    {
+      "id": 21,
+      "type": "GetImageSizeAndCount",
+      "pos": [
+        393,
+        282
+      ],
+      "size": {
+        "0": 210,
+        "1": 86
+      },
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 37
+        }
+      ],
+      "outputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "links": [
+            58
+          ],
+          "shape": 3,
+          "slot_index": 0
+        },
+        {
+          "name": "1440 width",
+          "type": "INT",
+          "links": null,
+          "shape": 3
+        },
+        {
+          "name": "1024 height",
+          "type": "INT",
+          "links": null,
+          "shape": 3
+        },
+        {
+          "name": "16 count",
+          "type": "INT",
+          "links": null,
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "GetImageSizeAndCount"
+      }
+    },
+    {
+      "id": 7,
+      "type": "Sam2Segmentation",
+      "pos": [
+        744,
+        227
+      ],
+      "size": {
+        "0": 314.3733825683594,
+        "1": 190.31338500976562
+      },
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "sam2_model",
+          "type": "SAM2MODEL",
+          "link": 9
+        },
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 58
+        },
+        {
+          "name": "bboxes",
+          "type": "BBOX",
+          "link": 54
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": null,
+          "slot_index": 3
+        },
+        {
+          "name": "coordinates_positive",
+          "type": "STRING",
+          "link": null,
+          "widget": {
+            "name": "coordinates_positive"
+          }
+        },
+        {
+          "name": "coordinates_negative",
+          "type": "STRING",
+          "link": null,
+          "widget": {
+            "name": "coordinates_negative"
+          }
+        }
+      ],
+      "outputs": [
+        {
+          "name": "mask",
+          "type": "MASK",
+          "links": [
+            43,
+            45
+          ],
+          "shape": 3,
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "Sam2Segmentation"
+      },
+      "widgets_values": [
+        true,
+        "",
+        "",
+        true
+      ]
+    },
+    {
+      "id": 30,
+      "type": "VHS_VideoCombine",
+      "pos": [
+        1093,
+        -868
+      ],
+      "size": [
+        772.7946166992188,
+        859.3206163194444
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 57,
+          "slot_index": 0
+        },
+        {
+          "name": "audio",
+          "type": "VHS_AUDIO",
+          "link": null
+        },
+        {
+          "name": "meta_batch",
+          "type": "VHS_BatchManager",
+          "link": null
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "Filenames",
+          "type": "VHS_FILENAMES",
+          "links": null,
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VHS_VideoCombine"
+      },
+      "widgets_values": {
+        "frame_rate": 3,
+        "loop_count": 0,
+        "filename_prefix": "AnimateDiff",
+        "format": "video/h264-mp4",
+        "pix_fmt": "yuv420p",
+        "crf": 19,
+        "save_metadata": true,
+        "pingpong": false,
+        "save_output": false,
+        "videopreview": {
+          "hidden": false,
+          "paused": false,
+          "params": {
+            "filename": "AnimateDiff_00001.mp4",
+            "subfolder": "",
+            "type": "temp",
+            "format": "video/h264-mp4",
+            "frame_rate": 3
+          }
+        }
+      }
+    },
+    {
+      "id": 16,
+      "type": "Florence2toCoordinates",
+      "pos": [
+        942,
+        16
+      ],
+      "size": {
+        "0": 315,
+        "1": 102
+      },
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "data",
+          "type": "JSON",
+          "link": 26
+        }
+      ],
+      "outputs": [
+        {
+          "name": "center_coordinates",
+          "type": "STRING",
+          "links": [],
+          "shape": 3,
+          "slot_index": 0
+        },
+        {
+          "name": "bboxes",
+          "type": "BBOX",
+          "links": [
+            54
+          ],
+          "shape": 3,
+          "slot_index": 1
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "Florence2toCoordinates"
+      },
+      "widgets_values": [
+        "0",
+        true
+      ]
+    },
+    {
+      "id": 12,
+      "type": "Florence2Run",
+      "pos": [
+        506,
+        -316
+      ],
+      "size": {
+        "0": 400,
+        "1": 352
+      },
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 28,
+          "slot_index": 0
+        },
+        {
+          "name": "florence2_model",
+          "type": "FL2MODEL",
+          "link": 23,
+          "slot_index": 1
+        }
+      ],
+      "outputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "links": [
+            55,
+            57
+          ],
+          "shape": 3,
+          "slot_index": 0
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "links": null,
+          "shape": 3
+        },
+        {
+          "name": "caption",
+          "type": "STRING",
+          "links": null,
+          "shape": 3
+        },
+        {
+          "name": "data",
+          "type": "JSON",
+          "links": [
+            26
+          ],
+          "shape": 3,
+          "slot_index": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "Florence2Run"
+      },
+      "widgets_values": [
+        "sword",
+        "caption_to_phrase_grounding",
+        true,
+        false,
+        1024,
+        3,
+        true,
+        "",
+        3228786869,
+        "fixed"
+      ]
+    }
+  ],
+  "links": [
+    [
+      9,
+      2,
+      0,
+      7,
+      0,
+      "SAM2MODEL"
+    ],
+    [
+      23,
+      13,
+      0,
+      12,
+      1,
+      "FL2MODEL"
+    ],
+    [
+      26,
+      12,
+      3,
+      16,
+      0,
+      "JSON"
+    ],
+    [
+      28,
+      11,
+      0,
+      12,
+      0,
+      "IMAGE"
+    ],
+    [
+      37,
+      11,
+      0,
+      21,
+      0,
+      "IMAGE"
+    ],
+    [
+      43,
+      7,
+      0,
+      26,
+      0,
+      "MASK"
+    ],
+    [
+      44,
+      26,
+      0,
+      25,
+      1,
+      "IMAGE"
+    ],
+    [
+      45,
+      7,
+      0,
+      25,
+      2,
+      "MASK"
+    ],
+    [
+      54,
+      16,
+      1,
+      7,
+      2,
+      "BBOX"
+    ],
+    [
+      55,
+      12,
+      0,
+      25,
+      0,
+      "IMAGE"
+    ],
+    [
+      56,
+      25,
+      0,
+      29,
+      0,
+      "IMAGE"
+    ],
+    [
+      57,
+      12,
+      0,
+      30,
+      0,
+      "IMAGE"
+    ],
+    [
+      58,
+      21,
+      0,
+      7,
+      1,
+      "IMAGE"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 0.620921323059155,
+      "offset": [
+        253.5867105157341,
+        622.007731477587
+      ]
+    }
+  },
+  "version": 0.4
+}

custom_nodes/comfyui-segment-anything-2/example_workflows/points_segment_video_example.json

@@ -0,0 +1,447 @@
+{
+  "last_node_id": 114,
+  "last_link_id": 54,
+  "nodes": [
+    {
+      "id": 106,
+      "type": "DownloadAndLoadSAM2Model",
+      "pos": [
+        451,
+        420
+      ],
+      "size": {
+        "0": 315,
+        "1": 130
+      },
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "outputs": [
+        {
+          "name": "sam2_model",
+          "type": "SAM2MODEL",
+          "links": [
+            40
+          ],
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DownloadAndLoadSAM2Model"
+      },
+      "widgets_values": [
+        "sam2_hiera_base_plus.safetensors",
+        "video",
+        "cuda",
+        "bf16"
+      ]
+    },
+    {
+      "id": 112,
+      "type": "ShowText|pysssss",
+      "pos": [
+        1421,
+        -359
+      ],
+      "size": {
+        "0": 315,
+        "1": 100
+      },
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "text",
+          "type": "STRING",
+          "link": 53,
+          "widget": {
+            "name": "text"
+          }
+        }
+      ],
+      "outputs": [
+        {
+          "name": "STRING",
+          "type": "STRING",
+          "links": null,
+          "shape": 6
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ShowText|pysssss"
+      },
+      "widgets_values": [
+        "",
+        "[{\"x\": 620, \"y\": 359}, {\"x\": 621, \"y\": 246}]"
+      ]
+    },
+    {
+      "id": 102,
+      "type": "VHS_LoadVideo",
+      "pos": [
+        14,
+        -59
+      ],
+      "size": [
+        363.24957275390625,
+        619.2495727539062
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "meta_batch",
+          "type": "VHS_BatchManager",
+          "link": null
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            41,
+            43,
+            52
+          ],
+          "shape": 3,
+          "slot_index": 0
+        },
+        {
+          "name": "frame_count",
+          "type": "INT",
+          "links": null,
+          "shape": 3
+        },
+        {
+          "name": "audio",
+          "type": "VHS_AUDIO",
+          "links": null,
+          "shape": 3
+        },
+        {
+          "name": "video_info",
+          "type": "VHS_VIDEOINFO",
+          "links": null,
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VHS_LoadVideo"
+      },
+      "widgets_values": {
+        "video": "ballerina_davinci.mp4",
+        "force_rate": 0,
+        "force_size": "Disabled",
+        "custom_width": 512,
+        "custom_height": 512,
+        "frame_load_cap": 16,
+        "skip_first_frames": 0,
+        "select_every_nth": 3,
+        "choose video to upload": "image",
+        "videopreview": {
+          "hidden": false,
+          "paused": false,
+          "params": {
+            "frame_load_cap": 16,
+            "skip_first_frames": 0,
+            "force_rate": 0,
+            "filename": "ballerina_davinci.mp4",
+            "type": "input",
+            "format": "video/mp4",
+            "select_every_nth": 3
+          }
+        }
+      }
+    },
+    {
+      "id": 113,
+      "type": "Note",
+      "pos": [
+        56,
+        -415
+      ],
+      "size": [
+        309.10654388427724,
+        177.01340377807628
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "properties": {
+        "text": ""
+      },
+      "widgets_values": [
+        "To get the image for the points editor, first create a canvas, then either input image/video (first frame is taken), or copy/paste an image while the node is selected, or drag&drop an image.\n\nWARNING: the image WILL BE SAVED to the node in compressed format, including when saving the workflow!\n\nClick the ? on the node for more information"
+      ],
+      "color": "#432",
+      "bgcolor": "#653"
+    },
+    {
+      "id": 107,
+      "type": "PreviewAnimation",
+      "pos": [
+        1340,
+        32
+      ],
+      "size": {
+        "0": 514.92431640625,
+        "1": 577.3973999023438
+      },
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 43
+        },
+        {
+          "name": "masks",
+          "type": "MASK",
+          "link": 42,
+          "slot_index": 1
+        }
+      ],
+      "title": "Preview Animation 16x768x768",
+      "properties": {
+        "Node name for S&R": "PreviewAnimation"
+      },
+      "widgets_values": [
+        16,
+        null
+      ]
+    },
+    {
+      "id": 105,
+      "type": "Sam2Segmentation",
+      "pos": [
+        859,
+        409
+      ],
+      "size": {
+        "0": 315,
+        "1": 170
+      },
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "sam2_model",
+          "type": "SAM2MODEL",
+          "link": 40,
+          "slot_index": 0
+        },
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 41,
+          "slot_index": 1
+        },
+        {
+          "name": "bboxes",
+          "type": "BBOX",
+          "link": null
+        },
+        {
+          "name": "coordinates_positive",
+          "type": "STRING",
+          "link": 54,
+          "widget": {
+            "name": "coordinates_positive"
+          },
+          "slot_index": 3
+        },
+        {
+          "name": "coordinates_negative",
+          "type": "STRING",
+          "link": null,
+          "widget": {
+            "name": "coordinates_negative"
+          }
+        }
+      ],
+      "outputs": [
+        {
+          "name": "mask",
+          "type": "MASK",
+          "links": [
+            42
+          ],
+          "shape": 3,
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "Sam2Segmentation"
+      },
+      "widgets_values": [
+        true,
+        "",
+        "",
+        false
+      ]
+    },
+    {
+      "id": 114,
+      "type": "PointsEditor",
+      "pos": [
+        432,
+        -735
+      ],
+      "size": [
+        813,
+        1068
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "bg_image",
+          "type": "IMAGE",
+          "link": 52
+        }
+      ],
+      "outputs": [
+        {
+          "name": "positive_coords",
+          "type": "STRING",
+          "links": [
+            53,
+            54
+          ],
+          "shape": 3,
+          "slot_index": 0
+        },
+        {
+          "name": "negative_coords",
+          "type": "STRING",
+          "links": null,
+          "shape": 3
+        },
+        {
+          "name": "bbox",
+          "type": "BBOX",
+          "links": null,
+          "shape": 3,
+          "slot_index": 2
+        },
+        {
+          "name": "bbox_mask",
+          "type": "MASK",
+          "links": null,
+          "shape": 3
+        },
+        {
+          "name": "cropped_image",
+          "type": "IMAGE",
+          "links": null,
+          "shape": 3
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "PointsEditor",
+        "imgData": {
+          "name": "bg_image",
+          "base64": [
+            "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"
+          ]
+        }
+      },
+      "widgets_values": [
+        "{\"positive\":[{\"x\":620.2460000000001,\"y\":359.37000000000006},{\"x\":620.73,\"y\":245.63000000000002}],\"negative\":[{\"x\":0,\"y\":0}]}",
+        "[{\"x\":620.2460000000001,\"y\":359.37000000000006},{\"x\":620.73,\"y\":245.63000000000002}]",
+        "[{\"x\":0,\"y\":0}]",
+        "[{}]",
+        "[{}]",
+        "xyxy",
+        768,
+        768,
+        false,
+        null,
+        null,
+        null
+      ]
+    }
+  ],
+  "links": [
+    [
+      40,
+      106,
+      0,
+      105,
+      0,
+      "SAM2MODEL"
+    ],
+    [
+      41,
+      102,
+      0,
+      105,
+      1,
+      "IMAGE"
+    ],
+    [
+      42,
+      105,
+      0,
+      107,
+      1,
+      "MASK"
+    ],
+    [
+      43,
+      102,
+      0,
+      107,
+      0,
+      "IMAGE"
+    ],
+    [
+      52,
+      102,
+      0,
+      114,
+      0,
+      "IMAGE"
+    ],
+    [
+      53,
+      114,
+      0,
+      112,
+      0,
+      "STRING"
+    ],
+    [
+      54,
+      114,
+      0,
+      105,
+      3,
+      "STRING"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 0.7513148009015777,
+      "offset": {
+        "0": 226.08052057760656,
+        "1": 820.3321624947772
+      }
+    }
+  },
+  "version": 0.4
+}

custom_nodes/comfyui-segment-anything-2/load_model.py

@@ -0,0 +1,194 @@
+import yaml
+from .sam2.modeling.sam2_base import SAM2Base
+from .sam2.modeling.backbones.image_encoder import ImageEncoder
+from .sam2.modeling.backbones.hieradet import Hiera
+from .sam2.modeling.backbones.image_encoder import FpnNeck
+from .sam2.modeling.position_encoding import PositionEmbeddingSine
+from .sam2.modeling.memory_attention import MemoryAttention, MemoryAttentionLayer
+from .sam2.modeling.sam.transformer import RoPEAttention
+from .sam2.modeling.memory_encoder import MemoryEncoder, MaskDownSampler, Fuser, CXBlock
+
+from .sam2.sam2_image_predictor import SAM2ImagePredictor
+from .sam2.sam2_video_predictor import SAM2VideoPredictor
+from .sam2.automatic_mask_generator import SAM2AutomaticMaskGenerator
+from comfy.utils import load_torch_file
+
+def load_model(model_path, model_cfg_path, segmentor, dtype, device):
+    # Load the YAML configuration
+    with open(model_cfg_path, 'r') as file:
+        config = yaml.safe_load(file)
+
+    # Extract the model configuration
+    model_config = config['model']
+
+    # Instantiate the image encoder components
+    trunk_config = model_config['image_encoder']['trunk']
+    neck_config = model_config['image_encoder']['neck']
+    position_encoding_config = neck_config['position_encoding']
+
+    position_encoding = PositionEmbeddingSine(
+        num_pos_feats=position_encoding_config['num_pos_feats'],
+        normalize=position_encoding_config['normalize'],
+        scale=position_encoding_config['scale'],
+        temperature=position_encoding_config['temperature']
+    )
+
+    neck = FpnNeck(
+        position_encoding=position_encoding,
+        d_model=neck_config['d_model'],
+        backbone_channel_list=neck_config['backbone_channel_list'],
+        fpn_top_down_levels=neck_config['fpn_top_down_levels'],
+        fpn_interp_model=neck_config['fpn_interp_model']
+    )
+
+    keys_to_include = ['embed_dim', 'num_heads', 'global_att_blocks', 'window_pos_embed_bkg_spatial_size', 'stages']
+    trunk_kwargs = {key: trunk_config[key] for key in keys_to_include if key in trunk_config}
+    trunk = Hiera(**trunk_kwargs)
+
+    image_encoder = ImageEncoder(
+        scalp=model_config['image_encoder']['scalp'],
+        trunk=trunk,
+        neck=neck
+    )
+    # Instantiate the memory attention components
+    memory_attention_layer_config = config['model']['memory_attention']['layer']
+    self_attention_config = memory_attention_layer_config['self_attention']
+    cross_attention_config = memory_attention_layer_config['cross_attention']
+
+    self_attention = RoPEAttention(
+        rope_theta=self_attention_config['rope_theta'],
+        feat_sizes=self_attention_config['feat_sizes'],
+        embedding_dim=self_attention_config['embedding_dim'],
+        num_heads=self_attention_config['num_heads'],
+        downsample_rate=self_attention_config['downsample_rate'],
+        dropout=self_attention_config['dropout']
+    )
+
+    cross_attention = RoPEAttention(
+        rope_theta=cross_attention_config['rope_theta'],
+        feat_sizes=cross_attention_config['feat_sizes'],
+        rope_k_repeat=cross_attention_config['rope_k_repeat'],
+        embedding_dim=cross_attention_config['embedding_dim'],
+        num_heads=cross_attention_config['num_heads'],
+        downsample_rate=cross_attention_config['downsample_rate'],
+        dropout=cross_attention_config['dropout'],
+        kv_in_dim=cross_attention_config['kv_in_dim']
+    )
+
+    memory_attention_layer = MemoryAttentionLayer(
+        activation=memory_attention_layer_config['activation'],
+        dim_feedforward=memory_attention_layer_config['dim_feedforward'],
+        dropout=memory_attention_layer_config['dropout'],
+        pos_enc_at_attn=memory_attention_layer_config['pos_enc_at_attn'],
+        self_attention=self_attention,
+        d_model=memory_attention_layer_config['d_model'],
+        pos_enc_at_cross_attn_keys=memory_attention_layer_config['pos_enc_at_cross_attn_keys'],
+        pos_enc_at_cross_attn_queries=memory_attention_layer_config['pos_enc_at_cross_attn_queries'],
+        cross_attention=cross_attention
+    )
+
+    memory_attention = MemoryAttention(
+        d_model=config['model']['memory_attention']['d_model'],
+        pos_enc_at_input=config['model']['memory_attention']['pos_enc_at_input'],
+        layer=memory_attention_layer,
+        num_layers=config['model']['memory_attention']['num_layers']
+    )
+
+    # Instantiate the memory encoder components
+    memory_encoder_config = config['model']['memory_encoder']
+    position_encoding_mem_enc_config = memory_encoder_config['position_encoding']
+    mask_downsampler_config = memory_encoder_config['mask_downsampler']
+    fuser_layer_config = memory_encoder_config['fuser']['layer']
+
+    position_encoding_mem_enc = PositionEmbeddingSine(
+        num_pos_feats=position_encoding_mem_enc_config['num_pos_feats'],
+        normalize=position_encoding_mem_enc_config['normalize'],
+        scale=position_encoding_mem_enc_config['scale'],
+        temperature=position_encoding_mem_enc_config['temperature']
+    )
+
+    mask_downsampler = MaskDownSampler(
+        kernel_size=mask_downsampler_config['kernel_size'],
+        stride=mask_downsampler_config['stride'],
+        padding=mask_downsampler_config['padding']
+    )
+
+    fuser_layer = CXBlock(
+        dim=fuser_layer_config['dim'],
+        kernel_size=fuser_layer_config['kernel_size'],
+        padding=fuser_layer_config['padding'],
+        layer_scale_init_value=float(fuser_layer_config['layer_scale_init_value'])
+    )
+    fuser = Fuser(
+        num_layers=memory_encoder_config['fuser']['num_layers'],
+        layer=fuser_layer
+    )
+
+    memory_encoder = MemoryEncoder(
+        position_encoding=position_encoding_mem_enc,
+        mask_downsampler=mask_downsampler,
+        fuser=fuser,
+        out_dim=memory_encoder_config['out_dim']
+    )
+
+    sam_mask_decoder_extra_args = {
+        "dynamic_multimask_via_stability": True,
+        "dynamic_multimask_stability_delta": 0.05,
+        "dynamic_multimask_stability_thresh": 0.98,
+    }
+
+    def initialize_model(model_class, model_config, segmentor, image_encoder, memory_attention, memory_encoder, sam_mask_decoder_extra_args, dtype, device):
+        return model_class(
+            image_encoder=image_encoder,
+            memory_attention=memory_attention,
+            memory_encoder=memory_encoder,
+            sam_mask_decoder_extra_args=sam_mask_decoder_extra_args,
+            num_maskmem=model_config['num_maskmem'],
+            image_size=model_config['image_size'],
+            sigmoid_scale_for_mem_enc=model_config['sigmoid_scale_for_mem_enc'],
+            sigmoid_bias_for_mem_enc=model_config['sigmoid_bias_for_mem_enc'],
+            use_mask_input_as_output_without_sam=model_config['use_mask_input_as_output_without_sam'],
+            directly_add_no_mem_embed=model_config['directly_add_no_mem_embed'],
+            use_high_res_features_in_sam=model_config['use_high_res_features_in_sam'],
+            multimask_output_in_sam=model_config['multimask_output_in_sam'],
+            iou_prediction_use_sigmoid=model_config['iou_prediction_use_sigmoid'],
+            use_obj_ptrs_in_encoder=model_config['use_obj_ptrs_in_encoder'],
+            add_tpos_enc_to_obj_ptrs=model_config['add_tpos_enc_to_obj_ptrs'],
+            only_obj_ptrs_in_the_past_for_eval=model_config['only_obj_ptrs_in_the_past_for_eval'],
+            pred_obj_scores=model_config['pred_obj_scores'],
+            pred_obj_scores_mlp=model_config['pred_obj_scores_mlp'],
+            fixed_no_obj_ptr=model_config['fixed_no_obj_ptr'],
+            multimask_output_for_tracking=model_config['multimask_output_for_tracking'],
+            use_multimask_token_for_obj_ptr=model_config['use_multimask_token_for_obj_ptr'],
+            compile_image_encoder=model_config['compile_image_encoder'],
+            multimask_min_pt_num=model_config['multimask_min_pt_num'],
+            multimask_max_pt_num=model_config['multimask_max_pt_num'],
+            use_mlp_for_obj_ptr_proj=model_config['use_mlp_for_obj_ptr_proj'],
+            proj_tpos_enc_in_obj_ptrs=model_config['proj_tpos_enc_in_obj_ptrs'],
+            no_obj_embed_spatial=model_config['no_obj_embed_spatial'],
+            use_signed_tpos_enc_to_obj_ptrs=model_config['use_signed_tpos_enc_to_obj_ptrs'],
+            binarize_mask_from_pts_for_mem_enc=True if segmentor == 'video' else False,
+        ).to(dtype).to(device).eval()
+
+    # Load the state dictionary
+    sd = load_torch_file(model_path)
+
+    # Initialize model based on segmentor type
+    if segmentor == 'single_image':
+        model_class = SAM2Base
+        model = initialize_model(model_class, model_config, segmentor, image_encoder, memory_attention, memory_encoder, sam_mask_decoder_extra_args, dtype, device)
+        model.load_state_dict(sd)
+        model = SAM2ImagePredictor(model)
+    elif segmentor == 'video':
+        model_class = SAM2VideoPredictor
+        model = initialize_model(model_class, model_config, segmentor, image_encoder, memory_attention, memory_encoder, sam_mask_decoder_extra_args, dtype, device)
+        model.load_state_dict(sd)
+    elif segmentor == 'automaskgenerator':
+        model_class = SAM2Base
+        model = initialize_model(model_class, model_config, segmentor, image_encoder, memory_attention, memory_encoder, sam_mask_decoder_extra_args, dtype, device)
+        model.load_state_dict(sd)
+        model = SAM2AutomaticMaskGenerator(model)
+    else:
+        raise ValueError(f"Segmentor {segmentor} not supported")
+
+    return model

custom_nodes/comfyui-segment-anything-2/nodes.py

@@ -0,0 +1,771 @@
+import torch
+from torch.functional import F
+import os
+import numpy as np
+import json
+import random
+
+from tqdm import tqdm
+from contextlib import nullcontext
+
+from .load_model import load_model
+
+import comfy.model_management as mm
+from comfy.utils import ProgressBar, common_upscale
+import folder_paths
+
+script_directory = os.path.dirname(os.path.abspath(__file__))
+
+class DownloadAndLoadSAM2Model:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "model": ([ 
+                    'sam2_hiera_base_plus.safetensors',
+                    'sam2_hiera_large.safetensors',
+                    'sam2_hiera_small.safetensors',
+                    'sam2_hiera_tiny.safetensors',
+                    'sam2.1_hiera_base_plus.safetensors',
+                    'sam2.1_hiera_large.safetensors',
+                    'sam2.1_hiera_small.safetensors',
+                    'sam2.1_hiera_tiny.safetensors',
+                    ],),
+            "segmentor": (
+                    ['single_image','video', 'automaskgenerator'],
+                    ),
+            "device": (['cuda', 'cpu', 'mps'], ),
+            "precision": ([ 'fp16','bf16','fp32'],
+                    {
+                    "default": 'fp16'
+                    }),
+
+            },
+        }
+
+    RETURN_TYPES = ("SAM2MODEL",)
+    RETURN_NAMES = ("sam2_model",)
+    FUNCTION = "loadmodel"
+    CATEGORY = "SAM2"
+
+    def loadmodel(self, model, segmentor, device, precision):
+        if precision != 'fp32' and device == 'cpu':
+            raise ValueError("fp16 and bf16 are not supported on cpu")
+
+        if device == "cuda":
+            if torch.cuda.get_device_properties(0).major >= 8:
+                # turn on tfloat32 for Ampere GPUs (https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices)
+                torch.backends.cuda.matmul.allow_tf32 = True
+                torch.backends.cudnn.allow_tf32 = True
+        dtype = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[precision]
+        device = {"cuda": torch.device("cuda"), "cpu": torch.device("cpu"), "mps": torch.device("mps")}[device]
+
+        download_path = os.path.join(folder_paths.models_dir, "sam2")
+        if precision != 'fp32' and "2.1" in model:
+            base_name, extension = model.rsplit('.', 1)
+            model = f"{base_name}-fp16.{extension}"
+        model_path = os.path.join(download_path, model)
+        print("model_path: ", model_path)
+        
+        if not os.path.exists(model_path):
+            print(f"Downloading SAM2 model to: {model_path}")
+            from huggingface_hub import snapshot_download
+            snapshot_download(repo_id="Kijai/sam2-safetensors",
+                            allow_patterns=[f"*{model}*"],
+                            local_dir=download_path,
+                            local_dir_use_symlinks=False)
+
+        model_mapping = {
+            "2.0": {
+                "base": "sam2_hiera_b+.yaml",
+                "large": "sam2_hiera_l.yaml",
+                "small": "sam2_hiera_s.yaml",
+                "tiny": "sam2_hiera_t.yaml"
+            },
+            "2.1": {
+                "base": "sam2.1_hiera_b+.yaml",
+                "large": "sam2.1_hiera_l.yaml",
+                "small": "sam2.1_hiera_s.yaml",
+                "tiny": "sam2.1_hiera_t.yaml"
+            }
+        }
+        version = "2.1" if "2.1" in model else "2.0"
+
+        model_cfg_path = next(
+            (os.path.join(script_directory, "sam2_configs", cfg) 
+            for key, cfg in model_mapping[version].items() if key in model),
+            None
+        )
+        print(f"Using model config: {model_cfg_path}")
+
+        model = load_model(model_path, model_cfg_path, segmentor, dtype, device)
+        
+        sam2_model = {
+            'model': model, 
+            'dtype': dtype,
+            'device': device,
+            'segmentor' : segmentor,
+            'version': version
+            }
+
+        return (sam2_model,)
+
+
+class Florence2toCoordinates:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "data": ("JSON", ),
+                "index": ("STRING", {"default": "0"}),
+                "batch": ("BOOLEAN", {"default": False}),
+            },
+            
+        }
+    
+    RETURN_TYPES = ("STRING", "BBOX")
+    RETURN_NAMES =("center_coordinates", "bboxes")
+    FUNCTION = "segment"
+    CATEGORY = "SAM2"
+
+    def segment(self, data, index, batch=False):
+        try:
+            coordinates = coordinates.replace("'", '"')
+            coordinates = json.loads(coordinates)
+        except:
+            coordinates = data
+
+        if len(data)==0:
+            return (json.dumps([{'x': 0, 'y': 0}]),)
+        center_points = []
+
+        def get_bboxes(item):
+            return item["bboxes"] if isinstance(item, dict) else item
+
+        if index.strip():  # Check if index is not empty
+            indexes = [int(i) for i in index.split(",")]
+        else:  # If index is empty, use all indices from data[0]
+            indexes = list(range(len(get_bboxes(data[0]))))
+
+        print("Indexes:", indexes)
+        bboxes = []
+        
+        if batch:
+            for idx in indexes:
+                if 0 <= idx < len(get_bboxes(data[0])):
+                    for i in range(len(data)):
+                        bbox = get_bboxes(data[i])[idx]
+                        min_x, min_y, max_x, max_y = bbox
+                        center_x = int((min_x + max_x) / 2)
+                        center_y = int((min_y + max_y) / 2)
+                        center_points.append({"x": center_x, "y": center_y})
+                        bboxes.append(bbox)
+        else:
+            for idx in indexes:
+                if 0 <= idx < len(get_bboxes(data[0])):
+                    bbox = get_bboxes(data[0])[idx]
+                    min_x, min_y, max_x, max_y = bbox
+                    center_x = int((min_x + max_x) / 2)
+                    center_y = int((min_y + max_y) / 2)
+                    center_points.append({"x": center_x, "y": center_y})
+                    bboxes.append(bbox)
+                else:
+                    raise ValueError(f"There's nothing in index: {idx}")
+                
+        coordinates = json.dumps(center_points)
+        print("Coordinates:", coordinates)
+        return (coordinates, bboxes)
+    
+class Sam2Segmentation:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "sam2_model": ("SAM2MODEL", ),
+                "image": ("IMAGE", ),
+                "keep_model_loaded": ("BOOLEAN", {"default": False}),
+            },
+            "optional": {
+                "coordinates_positive": ("STRING", {"forceInput": True}),
+                "coordinates_negative": ("STRING", {"forceInput": True}),
+                "bboxes": ("BBOX", ),
+                "individual_objects": ("BOOLEAN", {"default": False}),
+                "mask": ("MASK", ),
+                
+            },
+        }
+    
+    RETURN_TYPES = ("MASK", )
+    RETURN_NAMES =("mask", )
+    FUNCTION = "segment"
+    CATEGORY = "SAM2"
+
+    def segment(self, image, sam2_model, keep_model_loaded, coordinates_positive=None, coordinates_negative=None, 
+                individual_objects=False, bboxes=None, mask=None):
+        offload_device = mm.unet_offload_device()
+        model = sam2_model["model"]
+        device = sam2_model["device"]
+        dtype = sam2_model["dtype"]
+        segmentor = sam2_model["segmentor"]
+        B, H, W, C = image.shape
+        
+        if mask is not None:
+            input_mask = mask.clone().unsqueeze(1)
+            input_mask = F.interpolate(input_mask, size=(256, 256), mode="bilinear")
+            input_mask = input_mask.squeeze(1)
+
+        if segmentor == 'automaskgenerator':
+            raise ValueError("For automaskgenerator use Sam2AutoMaskSegmentation -node")
+        if segmentor == 'single_image' and B > 1:
+            print("Segmenting batch of images with single_image segmentor")
+
+        if segmentor == 'video' and bboxes is not None and "2.1" not in sam2_model["version"]:
+            raise ValueError("2.0 model doesn't support bboxes with video segmentor")
+
+        if segmentor == 'video': # video model needs images resized first thing
+            model_input_image_size = model.image_size
+            print("Resizing to model input image size: ", model_input_image_size)
+            image = common_upscale(image.movedim(-1,1), model_input_image_size, model_input_image_size, "bilinear", "disabled").movedim(1,-1)
+
+        #handle point coordinates
+        if coordinates_positive is not None:
+            try:
+                coordinates_positive = json.loads(coordinates_positive.replace("'", '"'))
+                coordinates_positive = [(coord['x'], coord['y']) for coord in coordinates_positive]
+                if coordinates_negative is not None:
+                    coordinates_negative = json.loads(coordinates_negative.replace("'", '"'))
+                    coordinates_negative = [(coord['x'], coord['y']) for coord in coordinates_negative]
+            except:
+                pass
+            
+            if not individual_objects:
+                positive_point_coords = np.atleast_2d(np.array(coordinates_positive))
+            else:
+                positive_point_coords = np.array([np.atleast_2d(coord) for coord in coordinates_positive])
+
+            if coordinates_negative is not None:
+                negative_point_coords = np.array(coordinates_negative)
+                # Ensure both positive and negative coords are lists of 2D arrays if individual_objects is True
+                if individual_objects:
+                    assert negative_point_coords.shape[0] <= positive_point_coords.shape[0], "Can't have more negative than positive points in individual_objects mode"
+                    if negative_point_coords.ndim == 2:
+                        negative_point_coords = negative_point_coords[:, np.newaxis, :]
+                    # Extend negative coordinates to match the number of positive coordinates
+                    while negative_point_coords.shape[0] < positive_point_coords.shape[0]:
+                        negative_point_coords = np.concatenate((negative_point_coords, negative_point_coords[:1, :, :]), axis=0)
+                    final_coords = np.concatenate((positive_point_coords, negative_point_coords), axis=1)
+                else:
+                    final_coords = np.concatenate((positive_point_coords, negative_point_coords), axis=0)
+            else:
+                final_coords = positive_point_coords
+
+        # Handle possible bboxes
+        if bboxes is not None:
+            boxes_np_batch = []
+            for bbox_list in bboxes:
+                boxes_np = []
+                for bbox in bbox_list:
+                    boxes_np.append(bbox)
+                boxes_np = np.array(boxes_np)
+                boxes_np_batch.append(boxes_np)
+            if individual_objects:
+                final_box = np.array(boxes_np_batch)
+            else:
+                final_box = np.array(boxes_np)
+            final_labels = None
+
+        #handle labels
+        if coordinates_positive is not None:
+            if not individual_objects:
+                positive_point_labels = np.ones(len(positive_point_coords))
+            else:
+                positive_labels = []
+                for point in positive_point_coords:
+                    positive_labels.append(np.array([1])) # 1)
+                positive_point_labels = np.stack(positive_labels, axis=0)
+                
+            if coordinates_negative is not None:
+                if not individual_objects:
+                    negative_point_labels = np.zeros(len(negative_point_coords))  # 0 = negative
+                    final_labels = np.concatenate((positive_point_labels, negative_point_labels), axis=0)
+                else:
+                    negative_labels = []
+                    for point in positive_point_coords:
+                        negative_labels.append(np.array([0])) # 1)
+                    negative_point_labels = np.stack(negative_labels, axis=0)
+                    #combine labels
+                    final_labels = np.concatenate((positive_point_labels, negative_point_labels), axis=1)                    
+            else:
+                final_labels = positive_point_labels
+            print("combined labels: ", final_labels)
+            print("combined labels shape: ", final_labels.shape)          
+        
+        mask_list = []
+        try:
+            model.to(device)
+        except:
+            model.model.to(device)
+        
+        autocast_condition = not mm.is_device_mps(device)
+        with torch.autocast(mm.get_autocast_device(device), dtype=dtype) if autocast_condition else nullcontext():
+            if segmentor == 'single_image':
+                image_np = (image.contiguous() * 255).byte().numpy()
+                comfy_pbar = ProgressBar(len(image_np))
+                tqdm_pbar = tqdm(total=len(image_np), desc="Processing Images")
+                for i in range(len(image_np)):
+                    model.set_image(image_np[i])
+                    if bboxes is None:
+                        input_box = None
+                    else:
+                        if len(image_np) > 1:
+                            input_box = final_box[i]
+                        input_box = final_box
+                    
+                    out_masks, scores, logits = model.predict(
+                        point_coords=final_coords if coordinates_positive is not None else None, 
+                        point_labels=final_labels if coordinates_positive is not None else None,
+                        box=input_box,
+                        multimask_output=True if not individual_objects else False,
+                        mask_input = input_mask[i].unsqueeze(0) if mask is not None else None,
+                        )
+                
+                    if out_masks.ndim == 3:
+                        sorted_ind = np.argsort(scores)[::-1]
+                        out_masks = out_masks[sorted_ind][0] #choose only the best result for now
+                        scores = scores[sorted_ind]
+                        logits = logits[sorted_ind]
+                        mask_list.append(np.expand_dims(out_masks, axis=0))
+                    else:
+                        _, _, H, W = out_masks.shape
+                        # Combine masks for all object IDs in the frame
+                        combined_mask = np.zeros((H, W), dtype=bool)
+                        for out_mask in out_masks:
+                            combined_mask = np.logical_or(combined_mask, out_mask)
+                        combined_mask = combined_mask.astype(np.uint8)
+                        mask_list.append(combined_mask)
+                    comfy_pbar.update(1)
+                    tqdm_pbar.update(1)
+
+            elif segmentor == 'video':
+                mask_list = []
+                if hasattr(self, 'inference_state') and self.inference_state is not None:
+                    model.reset_state(self.inference_state)
+                self.inference_state = model.init_state(image.permute(0, 3, 1, 2).contiguous(), H, W, device=device)
+                if bboxes is None:
+                        input_box = None
+                else:
+                    input_box = bboxes[0]
+                
+                if individual_objects and bboxes is not None:
+                    raise ValueError("bboxes not supported with individual_objects")
+
+
+                if individual_objects:
+                    for i, (coord, label) in enumerate(zip(final_coords, final_labels)):
+                        _, out_obj_ids, out_mask_logits = model.add_new_points_or_box(
+                        inference_state=self.inference_state,
+                        frame_idx=0,
+                        obj_id=i,
+                        points=final_coords[i],
+                        labels=final_labels[i],
+                        clear_old_points=True,
+                        box=input_box
+                        )
+                else:
+                    _, out_obj_ids, out_mask_logits = model.add_new_points_or_box(
+                        inference_state=self.inference_state,
+                        frame_idx=0,
+                        obj_id=1,
+                        points=final_coords if coordinates_positive is not None else None, 
+                        labels=final_labels if coordinates_positive is not None else None,
+                        clear_old_points=True,
+                        box=input_box
+                    )
+
+                pbar = ProgressBar(B)
+                video_segments = {}
+                for out_frame_idx, out_obj_ids, out_mask_logits in model.propagate_in_video(self.inference_state):
+                    video_segments[out_frame_idx] = {
+                        out_obj_id: (out_mask_logits[i] > 0.0).cpu().numpy()
+                        for i, out_obj_id in enumerate(out_obj_ids)
+                        }
+                    pbar.update(1)
+                    if individual_objects:
+                        _, _, H, W = out_mask_logits.shape
+                        # Combine masks for all object IDs in the frame
+                        combined_mask = np.zeros((H, W), dtype=np.uint8) 
+                        for i, out_obj_id in enumerate(out_obj_ids):
+                            out_mask = (out_mask_logits[i] > 0.0).cpu().numpy()
+                            combined_mask = np.logical_or(combined_mask, out_mask)
+                        video_segments[out_frame_idx] = combined_mask
+
+                if individual_objects:
+                    for frame_idx, combined_mask in video_segments.items():
+                        mask_list.append(combined_mask)
+                else:
+                    for frame_idx, obj_masks in video_segments.items():
+                        for out_obj_id, out_mask in obj_masks.items():
+                            mask_list.append(out_mask)
+
+        if not keep_model_loaded:
+            try:
+                model.to(offload_device)
+            except:
+                model.model.to(offload_device)
+            if hasattr(self, 'inference_state') and self.inference_state is not None and hasattr(model, "reset_state"):
+                model.reset_state(self.inference_state)
+                self.inference_state = None
+            mm.soft_empty_cache()
+        
+        out_list = []
+        for mask in mask_list:
+            mask_tensor = torch.from_numpy(mask)
+            mask_tensor = mask_tensor.permute(1, 2, 0)
+            mask_tensor = mask_tensor[:, :, 0]
+            out_list.append(mask_tensor)
+        mask_tensor = torch.stack(out_list, dim=0).cpu().float()
+        return (mask_tensor,)
+
+class Sam2VideoSegmentationAddPoints:
+    @classmethod
+    def IS_CHANGED(s): # TODO: smarter reset?
+        return ""
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "sam2_model": ("SAM2MODEL", ),
+                "coordinates_positive": ("STRING", {"forceInput": True}),
+                "frame_index": ("INT", {"default": 0}),
+                "object_index": ("INT", {"default": 0}),
+            },
+            "optional": {
+                "image": ("IMAGE", ),
+                "coordinates_negative": ("STRING", {"forceInput": True}),
+                "prev_inference_state": ("SAM2INFERENCESTATE", ),
+            },
+        }
+    
+    RETURN_TYPES = ("SAM2MODEL", "SAM2INFERENCESTATE", )
+    RETURN_NAMES =("sam2_model", "inference_state", )
+    FUNCTION = "segment"
+    CATEGORY = "SAM2"
+
+    def segment(self, sam2_model, coordinates_positive, frame_index, object_index, image=None, coordinates_negative=None, prev_inference_state=None):
+        offload_device = mm.unet_offload_device()
+        model = sam2_model["model"]
+        device = sam2_model["device"]
+        dtype = sam2_model["dtype"]
+        segmentor = sam2_model["segmentor"]
+        
+
+        if segmentor != 'video':
+            raise ValueError("Loaded model is not SAM2Video")
+        if image is not None:
+            B, H, W, C = image.shape
+            model_input_image_size = model.image_size
+            print("Resizing to model input image size: ", model_input_image_size)
+            image = common_upscale(image.movedim(-1,1), model_input_image_size, model_input_image_size, "bilinear", "disabled").movedim(1,-1)
+
+        try:
+            coordinates_positive = json.loads(coordinates_positive.replace("'", '"'))
+            coordinates_positive = [(coord['x'], coord['y']) for coord in coordinates_positive]
+            if coordinates_negative is not None:
+                coordinates_negative = json.loads(coordinates_negative.replace("'", '"'))
+                coordinates_negative = [(coord['x'], coord['y']) for coord in coordinates_negative]
+        except:
+            pass
+        
+        positive_point_coords = np.array(coordinates_positive)
+        positive_point_labels = [1] * len(positive_point_coords)  # 1 = positive
+        positive_point_labels = np.array(positive_point_labels)
+        print("positive coordinates: ", positive_point_coords)
+
+        if coordinates_negative is not None:
+            negative_point_coords = np.array(coordinates_negative)
+            negative_point_labels = [0] * len(negative_point_coords)  # 0 = negative
+            negative_point_labels = np.array(negative_point_labels)
+            print("negative coordinates: ", negative_point_coords)
+
+            # Combine coordinates and labels
+        else:
+            negative_point_coords = np.empty((0, 2))
+            negative_point_labels = np.array([])
+        # Ensure both positive and negative coordinates are 2D arrays
+        positive_point_coords = np.atleast_2d(positive_point_coords)
+        negative_point_coords = np.atleast_2d(negative_point_coords)
+
+        # Ensure both positive and negative labels are 1D arrays
+        positive_point_labels = np.atleast_1d(positive_point_labels)
+        negative_point_labels = np.atleast_1d(negative_point_labels)
+
+        combined_coords = np.concatenate((positive_point_coords, negative_point_coords), axis=0)
+        combined_labels = np.concatenate((positive_point_labels, negative_point_labels), axis=0)
+        
+        model.to(device)
+        
+        autocast_condition = not mm.is_device_mps(device)
+        with torch.autocast(mm.get_autocast_device(model.device), dtype=dtype) if autocast_condition else nullcontext(): 
+            if prev_inference_state is None:
+                print("Initializing inference state")
+                if hasattr(self, 'inference_state'):
+                    model.reset_state(self.inference_state)
+                self.inference_state = model.init_state(image.permute(0, 3, 1, 2).contiguous(), H, W, device=device)
+            else:
+                print("Using previous inference state")
+                B = prev_inference_state['num_frames']
+                self.inference_state = prev_inference_state['inference_state']
+            _, out_obj_ids, out_mask_logits = model.add_new_points(
+                inference_state=self.inference_state,
+                frame_idx=frame_index,
+                obj_id=object_index,
+                points=combined_coords,
+                labels=combined_labels,
+            )
+        inference_state = {
+            "inference_state": self.inference_state,
+            "num_frames": B,
+        }
+        sam2_model = {
+            'model': model, 
+            'dtype': dtype,
+            'device': device,
+            'segmentor' : segmentor
+            }    
+        return (sam2_model, inference_state,)
+
+class Sam2VideoSegmentation:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "sam2_model": ("SAM2MODEL", ),
+                "inference_state": ("SAM2INFERENCESTATE", ),
+                "keep_model_loaded": ("BOOLEAN", {"default": True}),
+            },
+        }
+    
+    RETURN_TYPES = ("MASK", )
+    RETURN_NAMES =("mask", )
+    FUNCTION = "segment"
+    CATEGORY = "SAM2"
+
+    def segment(self, sam2_model, inference_state, keep_model_loaded):
+        offload_device = mm.unet_offload_device()
+        model = sam2_model["model"]
+        device = sam2_model["device"]
+        dtype = sam2_model["dtype"]
+        segmentor = sam2_model["segmentor"]
+        inference_state = inference_state["inference_state"]
+        B = inference_state["num_frames"]
+
+        if segmentor != 'video':
+            raise ValueError("Loaded model is not SAM2Video")
+
+        model.to(device)
+        
+        autocast_condition = not mm.is_device_mps(device)
+        with torch.autocast(mm.get_autocast_device(device), dtype=dtype) if autocast_condition else nullcontext(): 
+            
+            #if hasattr(self, 'inference_state'):
+            #    model.reset_state(self.inference_state)
+
+            pbar = ProgressBar(B)
+            video_segments = {}
+            for out_frame_idx, out_obj_ids, out_mask_logits in model.propagate_in_video(inference_state):
+                print("out_mask_logits",out_mask_logits.shape)
+                _, _, H, W = out_mask_logits.shape
+                # Combine masks for all object IDs in the frame
+                combined_mask = np.zeros((H, W), dtype=np.uint8) 
+                for i, out_obj_id in enumerate(out_obj_ids):
+                    out_mask = (out_mask_logits[i] > 0.0).cpu().numpy()
+                    combined_mask = np.logical_or(combined_mask, out_mask)
+                video_segments[out_frame_idx] = combined_mask
+                pbar.update(1)
+
+            mask_list = []
+            # Collect the combined masks
+            for frame_idx, combined_mask in video_segments.items():
+                mask_list.append(combined_mask)
+            print(f"Total masks collected: {len(mask_list)}")
+
+        if not keep_model_loaded:
+            model.to(offload_device)
+        
+        out_list = []
+        for mask in mask_list:
+            mask_tensor = torch.from_numpy(mask)
+            mask_tensor = mask_tensor.permute(1, 2, 0)
+            mask_tensor = mask_tensor[:, :, 0]
+            out_list.append(mask_tensor)
+        mask_tensor = torch.stack(out_list, dim=0).cpu().float()
+        return (mask_tensor,)
+        
+class Sam2AutoSegmentation:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "sam2_model": ("SAM2MODEL", ),
+                "image": ("IMAGE", ),
+                "points_per_side": ("INT", {"default": 32}),
+                "points_per_batch": ("INT", {"default": 64}),
+                "pred_iou_thresh": ("FLOAT", {"default": 0.8, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "stability_score_thresh": ("FLOAT", {"default": 0.95, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "stability_score_offset": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "mask_threshold": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "crop_n_layers": ("INT", {"default": 0}),
+                "box_nms_thresh": ("FLOAT", {"default": 0.7, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "crop_nms_thresh": ("FLOAT", {"default": 0.7, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "crop_overlap_ratio": ("FLOAT", {"default": 0.34, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "crop_n_points_downscale_factor": ("INT", {"default": 1}),
+                "min_mask_region_area": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "use_m2m": ("BOOLEAN", {"default": False}),
+                "keep_model_loaded": ("BOOLEAN", {"default": True}),
+            },
+        }
+    
+    RETURN_TYPES = ("MASK", "IMAGE", "BBOX",)
+    RETURN_NAMES =("mask", "segmented_image", "bbox" ,)
+    FUNCTION = "segment"
+    CATEGORY = "SAM2"
+
+    def segment(self, image, sam2_model, points_per_side, points_per_batch, pred_iou_thresh, stability_score_thresh, 
+                stability_score_offset, crop_n_layers, box_nms_thresh, crop_n_points_downscale_factor, min_mask_region_area, 
+                use_m2m, mask_threshold, crop_nms_thresh, crop_overlap_ratio, keep_model_loaded):
+        offload_device = mm.unet_offload_device()
+        model = sam2_model["model"]
+        device = sam2_model["device"]
+        dtype = sam2_model["dtype"]
+        segmentor = sam2_model["segmentor"]
+        
+        if segmentor != 'automaskgenerator':
+            raise ValueError("Loaded model is not SAM2AutomaticMaskGenerator")
+        
+        model.points_per_side=points_per_side
+        model.points_per_batch=points_per_batch
+        model.pred_iou_thresh=pred_iou_thresh
+        model.stability_score_thresh=stability_score_thresh
+        model.stability_score_offset=stability_score_offset
+        model.crop_n_layers=crop_n_layers
+        model.box_nms_thresh=box_nms_thresh
+        model.crop_n_points_downscale_factor=crop_n_points_downscale_factor
+        model.crop_nms_thresh=crop_nms_thresh
+        model.crop_overlap_ratio=crop_overlap_ratio
+        model.min_mask_region_area=min_mask_region_area
+        model.use_m2m=use_m2m
+        model.mask_threshold=mask_threshold
+        
+        model.predictor.model.to(device)
+        
+        B, H, W, C = image.shape
+        image_np = (image.contiguous() * 255).byte().numpy()
+
+        out_list = []
+        segment_out_list = []
+        mask_list=[]
+        
+        pbar = ProgressBar(B)
+        autocast_condition = not mm.is_device_mps(device)
+        with torch.autocast(mm.get_autocast_device(device), dtype=dtype) if autocast_condition else nullcontext():
+            for img_np in image_np:
+                result_dict = model.generate(img_np)
+                mask_list = [item['segmentation'] for item in result_dict]
+                bbox_list = [item['bbox'] for item in result_dict]
+
+                # Generate random colors for each mask
+                num_masks = len(mask_list)
+                colors = [tuple(random.choices(range(256), k=3)) for _ in range(num_masks)]
+                
+                # Create a blank image to overlay masks
+                overlay_image = np.zeros((H, W, 3), dtype=np.uint8)
+
+                # Create a combined mask initialized to zeros
+                combined_mask = np.zeros((H, W), dtype=np.uint8)
+
+                # Iterate through masks and color them
+                for mask, color in zip(mask_list, colors):
+
+                    # Combine masks using logical OR
+                    combined_mask = np.logical_or(combined_mask, mask).astype(np.uint8)
+                    
+                    # Convert mask to numpy array
+                    mask_np = mask.astype(np.uint8)
+                    
+                    # Color the mask
+                    colored_mask = np.zeros_like(overlay_image)
+                    for i in range(3):  # Apply color channel-wise
+                        colored_mask[:, :, i] = mask_np * color[i]
+                    
+                    # Blend the colored mask with the overlay image
+                    overlay_image = np.where(colored_mask > 0, colored_mask, overlay_image)
+                out_list.append(torch.from_numpy(combined_mask))
+                segment_out_list.append(overlay_image)
+                pbar.update(1)
+
+        stacked_array = np.stack(segment_out_list, axis=0)
+        segment_image_tensor = torch.from_numpy(stacked_array).float() / 255
+
+        if not keep_model_loaded:
+           model.predictor.model.to(offload_device)
+        
+        mask_tensor = torch.stack(out_list, dim=0)
+        return (mask_tensor.cpu().float(), segment_image_tensor.cpu().float(), bbox_list)
+
+#WIP    
+# class OwlV2Detector:
+#     @classmethod
+#     def INPUT_TYPES(s):
+#         return {
+#             "required": {
+#                 "image": ("IMAGE", ),
+#             },
+#         }
+    
+#     RETURN_TYPES = ("MASK", )
+#     RETURN_NAMES =("mask", )
+#     FUNCTION = "segment"
+#     CATEGORY = "SAM2"
+
+#     def segment(self, image):
+#         from transformers import Owlv2Processor, Owlv2ForObjectDetection
+#         device = mm.get_torch_device()
+#         offload_device = mm.unet_offload_device()
+#         processor = Owlv2Processor.from_pretrained("google/owlv2-base-patch16-ensemble")
+#         model = Owlv2ForObjectDetection.from_pretrained("google/owlv2-base-patch16-ensemble")
+
+#         url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+#         image = Image.open(requests.get(url, stream=True).raw)
+#         texts = [["a photo of a cat", "a photo of a dog"]]
+#         inputs = processor(text=texts, images=image, return_tensors="pt")
+#         outputs = model(**inputs)
+
+#         # Target image sizes (height, width) to rescale box predictions [batch_size, 2]
+#         target_sizes = torch.Tensor([image.size[::-1]])
+#         # Convert outputs (bounding boxes and class logits) to Pascal VOC Format (xmin, ymin, xmax, ymax)
+#         results = processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes, threshold=0.1)
+#         i = 0  # Retrieve predictions for the first image for the corresponding text queries
+#         text = texts[i]
+#         boxes, scores, labels = results[i]["boxes"], results[i]["scores"], results[i]["labels"]
+#         for box, score, label in zip(boxes, scores, labels):
+#             box = [round(i, 2) for i in box.tolist()]
+#             print(f"Detected {text[label]} with confidence {round(score.item(), 3)} at location {box}")
+
+
+#         return (mask_tensor,)
+     
+NODE_CLASS_MAPPINGS = {
+    "DownloadAndLoadSAM2Model": DownloadAndLoadSAM2Model,
+    "Sam2Segmentation": Sam2Segmentation,
+    "Florence2toCoordinates": Florence2toCoordinates,
+    "Sam2AutoSegmentation": Sam2AutoSegmentation,
+    "Sam2VideoSegmentationAddPoints": Sam2VideoSegmentationAddPoints,
+    "Sam2VideoSegmentation": Sam2VideoSegmentation
+}
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "DownloadAndLoadSAM2Model": "(Down)Load SAM2Model",
+    "Sam2Segmentation": "Sam2Segmentation",
+    "Florence2toCoordinates": "Florence2 Coordinates",
+    "Sam2AutoSegmentation": "Sam2AutoSegmentation",
+    "Sam2VideoSegmentationAddPoints": "Sam2VideoSegmentationAddPoints",
+    "Sam2VideoSegmentation": "Sam2VideoSegmentation"
+}

custom_nodes/comfyui-segment-anything-2/pyproject.toml

@@ -0,0 +1,15 @@
+[project]
+name = "comfyui-segment-anything-2"
+description = "Nodes to use [a/segment-anything-2](https://github.com/facebookresearch/segment-anything-2) for image or video segmentation."
+version = "1.0.2"
+license = {file = "LICENSE"}
+dependencies = []
+
+[project.urls]
+Repository = "https://github.com/kijai/ComfyUI-segment-anything-2"
+#  Used by Comfy Registry https://comfyregistry.org
+
+[tool.comfy]
+PublisherId = "kijai"
+DisplayName = "ComfyUI-segment-anything-2"
+Icon = ""

custom_nodes/comfyui-segment-anything-2/readme.md

@@ -0,0 +1,25 @@
+# WORK IN PROGRESS
+
+PointsEditor is now available for testing in KJNodes: https://github.com/kijai/ComfyUI-KJNodes
+
+https://github.com/user-attachments/assets/c4a88647-679f-4cf2-ba1f-4fa8c7308c1e
+
+https://github.com/user-attachments/assets/f15fafe8-72e8-41cc-b246-e947b1efe5ec
+
+https://github.com/user-attachments/assets/c1efb595-0fb1-4ae7-b4fa-2def08eda0a8
+
+For testing only currently.
+
+Functional, but needs better coordinate selector. 
+
+For now mask postprocessing is disabled due to it needing cuda extension compilation. We can use other nodes for this purpose anyway, so might leave it that way, we'll see.
+
+Models are automatically downloade from https://huggingface.co/Kijai/sam2-safetensors/tree/main
+
+to `ComfyUI/models/sam2`
+
+
+
+Original repo:
+
+https://github.com/facebookresearch/segment-anything-2

custom_nodes/comfyui-segment-anything-2/sam2/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

custom_nodes/comfyui-segment-anything-2/sam2/automatic_mask_generator.py

@@ -0,0 +1,436 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Adapted from https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/automatic_mask_generator.py
+from typing import Any, Dict, List, Optional, Tuple
+
+import numpy as np
+import torch
+from torchvision.ops.boxes import batched_nms, box_area  # type: ignore
+
+from ..sam2.modeling.sam2_base import SAM2Base
+from ..sam2.sam2_image_predictor import SAM2ImagePredictor
+from ..sam2.utils.amg import (
+    area_from_rle,
+    batch_iterator,
+    batched_mask_to_box,
+    box_xyxy_to_xywh,
+    build_all_layer_point_grids,
+    calculate_stability_score,
+    coco_encode_rle,
+    generate_crop_boxes,
+    is_box_near_crop_edge,
+    mask_to_rle_pytorch,
+    MaskData,
+    remove_small_regions,
+    rle_to_mask,
+    uncrop_boxes_xyxy,
+    uncrop_masks,
+    uncrop_points,
+)
+
+
+class SAM2AutomaticMaskGenerator:
+    def __init__(
+        self,
+        model: SAM2Base,
+        points_per_side: Optional[int] = 32,
+        points_per_batch: int = 64,
+        pred_iou_thresh: float = 0.8,
+        stability_score_thresh: float = 0.95,
+        stability_score_offset: float = 1.0,
+        mask_threshold: float = 0.0,
+        box_nms_thresh: float = 0.7,
+        crop_n_layers: int = 0,
+        crop_nms_thresh: float = 0.7,
+        crop_overlap_ratio: float = 512 / 1500,
+        crop_n_points_downscale_factor: int = 1,
+        point_grids: Optional[List[np.ndarray]] = None,
+        min_mask_region_area: int = 0,
+        output_mode: str = "binary_mask",
+        use_m2m: bool = False,
+        multimask_output: bool = True,
+    ) -> None:
+        """
+        Using a SAM 2 model, generates masks for the entire image.
+        Generates a grid of point prompts over the image, then filters
+        low quality and duplicate masks. The default settings are chosen
+        for SAM 2 with a HieraL backbone.
+
+        Arguments:
+          model (Sam): The SAM 2 model to use for mask prediction.
+          points_per_side (int or None): The number of points to be sampled
+            along one side of the image. The total number of points is
+            points_per_side**2. If None, 'point_grids' must provide explicit
+            point sampling.
+          points_per_batch (int): Sets the number of points run simultaneously
+            by the model. Higher numbers may be faster but use more GPU memory.
+          pred_iou_thresh (float): A filtering threshold in [0,1], using the
+            model's predicted mask quality.
+          stability_score_thresh (float): A filtering threshold in [0,1], using
+            the stability of the mask under changes to the cutoff used to binarize
+            the model's mask predictions.
+          stability_score_offset (float): The amount to shift the cutoff when
+            calculated the stability score.
+          mask_threshold (float): Threshold for binarizing the mask logits
+          box_nms_thresh (float): The box IoU cutoff used by non-maximal
+            suppression to filter duplicate masks.
+          crop_n_layers (int): If >0, mask prediction will be run again on
+            crops of the image. Sets the number of layers to run, where each
+            layer has 2**i_layer number of image crops.
+          crop_nms_thresh (float): The box IoU cutoff used by non-maximal
+            suppression to filter duplicate masks between different crops.
+          crop_overlap_ratio (float): Sets the degree to which crops overlap.
+            In the first crop layer, crops will overlap by this fraction of
+            the image length. Later layers with more crops scale down this overlap.
+          crop_n_points_downscale_factor (int): The number of points-per-side
+            sampled in layer n is scaled down by crop_n_points_downscale_factor**n.
+          point_grids (list(np.ndarray) or None): A list over explicit grids
+            of points used for sampling, normalized to [0,1]. The nth grid in the
+            list is used in the nth crop layer. Exclusive with points_per_side.
+          min_mask_region_area (int): If >0, postprocessing will be applied
+            to remove disconnected regions and holes in masks with area smaller
+            than min_mask_region_area. Requires opencv.
+          output_mode (str): The form masks are returned in. Can be 'binary_mask',
+            'uncompressed_rle', or 'coco_rle'. 'coco_rle' requires pycocotools.
+            For large resolutions, 'binary_mask' may consume large amounts of
+            memory.
+          use_m2m (bool): Whether to add a one step refinement using previous mask predictions.
+          multimask_output (bool): Whether to output multimask at each point of the grid.
+        """
+
+        assert (points_per_side is None) != (
+            point_grids is None
+        ), "Exactly one of points_per_side or point_grid must be provided."
+        if points_per_side is not None:
+            self.point_grids = build_all_layer_point_grids(
+                points_per_side,
+                crop_n_layers,
+                crop_n_points_downscale_factor,
+            )
+        elif point_grids is not None:
+            self.point_grids = point_grids
+        else:
+            raise ValueError("Can't have both points_per_side and point_grid be None.")
+
+        assert output_mode in [
+            "binary_mask",
+            "uncompressed_rle",
+            "coco_rle",
+        ], f"Unknown output_mode {output_mode}."
+        if output_mode == "coco_rle":
+            try:
+                from pycocotools import mask as mask_utils  # type: ignore  # noqa: F401
+            except ImportError as e:
+                print("Please install pycocotools")
+                raise e
+
+        self.predictor = SAM2ImagePredictor(
+            model,
+            max_hole_area=min_mask_region_area,
+            max_sprinkle_area=min_mask_region_area,
+        )
+        self.points_per_batch = points_per_batch
+        self.pred_iou_thresh = pred_iou_thresh
+        self.stability_score_thresh = stability_score_thresh
+        self.stability_score_offset = stability_score_offset
+        self.mask_threshold = mask_threshold
+        self.box_nms_thresh = box_nms_thresh
+        self.crop_n_layers = crop_n_layers
+        self.crop_nms_thresh = crop_nms_thresh
+        self.crop_overlap_ratio = crop_overlap_ratio
+        self.crop_n_points_downscale_factor = crop_n_points_downscale_factor
+        self.min_mask_region_area = min_mask_region_area
+        self.output_mode = output_mode
+        self.use_m2m = use_m2m
+        self.multimask_output = multimask_output
+
+    @torch.no_grad()
+    def generate(self, image: np.ndarray) -> List[Dict[str, Any]]:
+        """
+        Generates masks for the given image.
+
+        Arguments:
+          image (np.ndarray): The image to generate masks for, in HWC uint8 format.
+
+        Returns:
+           list(dict(str, any)): A list over records for masks. Each record is
+             a dict containing the following keys:
+               segmentation (dict(str, any) or np.ndarray): The mask. If
+                 output_mode='binary_mask', is an array of shape HW. Otherwise,
+                 is a dictionary containing the RLE.
+               bbox (list(float)): The box around the mask, in XYWH format.
+               area (int): The area in pixels of the mask.
+               predicted_iou (float): The model's own prediction of the mask's
+                 quality. This is filtered by the pred_iou_thresh parameter.
+               point_coords (list(list(float))): The point coordinates input
+                 to the model to generate this mask.
+               stability_score (float): A measure of the mask's quality. This
+                 is filtered on using the stability_score_thresh parameter.
+               crop_box (list(float)): The crop of the image used to generate
+                 the mask, given in XYWH format.
+        """
+
+        # Generate masks
+        mask_data = self._generate_masks(image)
+
+        # Encode masks
+        if self.output_mode == "coco_rle":
+            mask_data["segmentations"] = [
+                coco_encode_rle(rle) for rle in mask_data["rles"]
+            ]
+        elif self.output_mode == "binary_mask":
+            mask_data["segmentations"] = [rle_to_mask(rle) for rle in mask_data["rles"]]
+        else:
+            mask_data["segmentations"] = mask_data["rles"]
+
+        # Write mask records
+        curr_anns = []
+        for idx in range(len(mask_data["segmentations"])):
+            ann = {
+                "segmentation": mask_data["segmentations"][idx],
+                "area": area_from_rle(mask_data["rles"][idx]),
+                "bbox": box_xyxy_to_xywh(mask_data["boxes"][idx]).tolist(),
+                "predicted_iou": mask_data["iou_preds"][idx].item(),
+                "point_coords": [mask_data["points"][idx].tolist()],
+                "stability_score": mask_data["stability_score"][idx].item(),
+                "crop_box": box_xyxy_to_xywh(mask_data["crop_boxes"][idx]).tolist(),
+            }
+            curr_anns.append(ann)
+
+        return curr_anns
+
+    def _generate_masks(self, image: np.ndarray) -> MaskData:
+        orig_size = image.shape[:2]
+        crop_boxes, layer_idxs = generate_crop_boxes(
+            orig_size, self.crop_n_layers, self.crop_overlap_ratio
+        )
+
+        # Iterate over image crops
+        data = MaskData()
+        for crop_box, layer_idx in zip(crop_boxes, layer_idxs):
+            crop_data = self._process_crop(image, crop_box, layer_idx, orig_size)
+            data.cat(crop_data)
+
+        # Remove duplicate masks between crops
+        if len(crop_boxes) > 1:
+            # Prefer masks from smaller crops
+            scores = 1 / box_area(data["crop_boxes"])
+            scores = scores.to(data["boxes"].device)
+            keep_by_nms = batched_nms(
+                data["boxes"].float(),
+                scores,
+                torch.zeros_like(data["boxes"][:, 0]),  # categories
+                iou_threshold=self.crop_nms_thresh,
+            )
+            data.filter(keep_by_nms)
+        data.to_numpy()
+        return data
+
+    def _process_crop(
+        self,
+        image: np.ndarray,
+        crop_box: List[int],
+        crop_layer_idx: int,
+        orig_size: Tuple[int, ...],
+    ) -> MaskData:
+        # Crop the image and calculate embeddings
+        x0, y0, x1, y1 = crop_box
+        cropped_im = image[y0:y1, x0:x1, :]
+        cropped_im_size = cropped_im.shape[:2]
+        self.predictor.set_image(cropped_im)
+
+        # Get points for this crop
+        points_scale = np.array(cropped_im_size)[None, ::-1]
+        points_for_image = self.point_grids[crop_layer_idx] * points_scale
+
+        # Generate masks for this crop in batches
+        data = MaskData()
+        for (points,) in batch_iterator(self.points_per_batch, points_for_image):
+            batch_data = self._process_batch(
+                points, cropped_im_size, crop_box, orig_size, normalize=True
+            )
+            data.cat(batch_data)
+            del batch_data
+        self.predictor.reset_predictor()
+
+        # Remove duplicates within this crop.
+        keep_by_nms = batched_nms(
+            data["boxes"].float(),
+            data["iou_preds"],
+            torch.zeros_like(data["boxes"][:, 0]),  # categories
+            iou_threshold=self.box_nms_thresh,
+        )
+        data.filter(keep_by_nms)
+
+        # Return to the original image frame
+        data["boxes"] = uncrop_boxes_xyxy(data["boxes"], crop_box)
+        data["points"] = uncrop_points(data["points"], crop_box)
+        data["crop_boxes"] = torch.tensor([crop_box for _ in range(len(data["rles"]))])
+
+        return data
+
+    def _process_batch(
+        self,
+        points: np.ndarray,
+        im_size: Tuple[int, ...],
+        crop_box: List[int],
+        orig_size: Tuple[int, ...],
+        normalize=False,
+    ) -> MaskData:
+        orig_h, orig_w = orig_size
+
+        # Run model on this batch
+        points = torch.as_tensor(
+            points, dtype=torch.float32, device=self.predictor.device
+        )
+        in_points = self.predictor._transforms.transform_coords(
+            points, normalize=normalize, orig_hw=im_size
+        )
+        in_labels = torch.ones(
+            in_points.shape[0], dtype=torch.int, device=in_points.device
+        )
+        masks, iou_preds, low_res_masks = self.predictor._predict(
+            in_points[:, None, :],
+            in_labels[:, None],
+            multimask_output=self.multimask_output,
+            return_logits=True,
+        )
+
+        # Serialize predictions and store in MaskData
+        data = MaskData(
+            masks=masks.flatten(0, 1),
+            iou_preds=iou_preds.flatten(0, 1),
+            points=points.repeat_interleave(masks.shape[1], dim=0),
+            low_res_masks=low_res_masks.flatten(0, 1),
+        )
+        del masks
+
+        if not self.use_m2m:
+            # Filter by predicted IoU
+            if self.pred_iou_thresh > 0.0:
+                keep_mask = data["iou_preds"] > self.pred_iou_thresh
+                data.filter(keep_mask)
+
+            # Calculate and filter by stability score
+            data["stability_score"] = calculate_stability_score(
+                data["masks"], self.mask_threshold, self.stability_score_offset
+            )
+            if self.stability_score_thresh > 0.0:
+                keep_mask = data["stability_score"] >= self.stability_score_thresh
+                data.filter(keep_mask)
+        else:
+            # One step refinement using previous mask predictions
+            in_points = self.predictor._transforms.transform_coords(
+                data["points"], normalize=normalize, orig_hw=im_size
+            )
+            labels = torch.ones(
+                in_points.shape[0], dtype=torch.int, device=in_points.device
+            )
+            masks, ious = self.refine_with_m2m(
+                in_points, labels, data["low_res_masks"], self.points_per_batch
+            )
+            data["masks"] = masks.squeeze(1)
+            data["iou_preds"] = ious.squeeze(1)
+
+            if self.pred_iou_thresh > 0.0:
+                keep_mask = data["iou_preds"] > self.pred_iou_thresh
+                data.filter(keep_mask)
+
+            data["stability_score"] = calculate_stability_score(
+                data["masks"], self.mask_threshold, self.stability_score_offset
+            )
+            if self.stability_score_thresh > 0.0:
+                keep_mask = data["stability_score"] >= self.stability_score_thresh
+                data.filter(keep_mask)
+
+        # Threshold masks and calculate boxes
+        data["masks"] = data["masks"] > self.mask_threshold
+        data["boxes"] = batched_mask_to_box(data["masks"])
+
+        # Filter boxes that touch crop boundaries
+        keep_mask = ~is_box_near_crop_edge(
+            data["boxes"], crop_box, [0, 0, orig_w, orig_h]
+        )
+        if not torch.all(keep_mask):
+            data.filter(keep_mask)
+
+        # Compress to RLE
+        data["masks"] = uncrop_masks(data["masks"], crop_box, orig_h, orig_w)
+        data["rles"] = mask_to_rle_pytorch(data["masks"])
+        del data["masks"]
+
+        return data
+
+    @staticmethod
+    def postprocess_small_regions(
+        mask_data: MaskData, min_area: int, nms_thresh: float
+    ) -> MaskData:
+        """
+        Removes small disconnected regions and holes in masks, then reruns
+        box NMS to remove any new duplicates.
+
+        Edits mask_data in place.
+
+        Requires open-cv as a dependency.
+        """
+        if len(mask_data["rles"]) == 0:
+            return mask_data
+
+        # Filter small disconnected regions and holes
+        new_masks = []
+        scores = []
+        for rle in mask_data["rles"]:
+            mask = rle_to_mask(rle)
+
+            mask, changed = remove_small_regions(mask, min_area, mode="holes")
+            unchanged = not changed
+            mask, changed = remove_small_regions(mask, min_area, mode="islands")
+            unchanged = unchanged and not changed
+
+            new_masks.append(torch.as_tensor(mask).unsqueeze(0))
+            # Give score=0 to changed masks and score=1 to unchanged masks
+            # so NMS will prefer ones that didn't need postprocessing
+            scores.append(float(unchanged))
+
+        # Recalculate boxes and remove any new duplicates
+        masks = torch.cat(new_masks, dim=0)
+        boxes = batched_mask_to_box(masks)
+        keep_by_nms = batched_nms(
+            boxes.float(),
+            torch.as_tensor(scores),
+            torch.zeros_like(boxes[:, 0]),  # categories
+            iou_threshold=nms_thresh,
+        )
+
+        # Only recalculate RLEs for masks that have changed
+        for i_mask in keep_by_nms:
+            if scores[i_mask] == 0.0:
+                mask_torch = masks[i_mask].unsqueeze(0)
+                mask_data["rles"][i_mask] = mask_to_rle_pytorch(mask_torch)[0]
+                mask_data["boxes"][i_mask] = boxes[i_mask]  # update res directly
+        mask_data.filter(keep_by_nms)
+
+        return mask_data
+
+    def refine_with_m2m(self, points, point_labels, low_res_masks, points_per_batch):
+        new_masks = []
+        new_iou_preds = []
+
+        for cur_points, cur_point_labels, low_res_mask in batch_iterator(
+            points_per_batch, points, point_labels, low_res_masks
+        ):
+            best_masks, best_iou_preds, _ = self.predictor._predict(
+                cur_points[:, None, :],
+                cur_point_labels[:, None],
+                mask_input=low_res_mask[:, None, :],
+                multimask_output=False,
+                return_logits=True,
+            )
+            new_masks.append(best_masks)
+            new_iou_preds.append(best_iou_preds)
+        masks = torch.cat(new_masks, dim=0)
+        return masks, torch.cat(new_iou_preds, dim=0)

custom_nodes/comfyui-segment-anything-2/sam2/modeling/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

custom_nodes/comfyui-segment-anything-2/sam2/modeling/backbones/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

custom_nodes/comfyui-segment-anything-2/sam2/modeling/backbones/hieradet.py

@@ -0,0 +1,316 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from functools import partial
+from typing import List, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+#from iopath.common.file_io import g_pathmgr
+
+from ....sam2.modeling.backbones.utils import (
+    PatchEmbed,
+    window_partition,
+    window_unpartition,
+)
+
+from ....sam2.modeling.sam2_utils import DropPath, MLP
+
+
+def do_pool(x: torch.Tensor, pool: nn.Module, norm: nn.Module = None) -> torch.Tensor:
+    if pool is None:
+        return x
+    # (B, H, W, C) -> (B, C, H, W)
+    x = x.permute(0, 3, 1, 2)
+    x = pool(x)
+    # (B, C, H', W') -> (B, H', W', C)
+    x = x.permute(0, 2, 3, 1)
+    if norm:
+        x = norm(x)
+
+    return x
+
+
+class MultiScaleAttention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        dim_out: int,
+        num_heads: int,
+        q_pool: nn.Module = None,
+    ):
+        super().__init__()
+
+        self.dim = dim
+        self.dim_out = dim_out
+        self.num_heads = num_heads
+        self.q_pool = q_pool
+        self.qkv = nn.Linear(dim, dim_out * 3)
+        self.proj = nn.Linear(dim_out, dim_out)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, H, W, _ = x.shape
+        # qkv with shape (B, H * W, 3, nHead, C)
+        qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1)
+        # q, k, v with shape (B, H * W, nheads, C)
+        q, k, v = torch.unbind(qkv, 2)
+
+        # Q pooling (for downsample at stage changes)
+        if self.q_pool:
+            q = do_pool(q.reshape(B, H, W, -1), self.q_pool)
+            H, W = q.shape[1:3]  # downsampled shape
+            q = q.reshape(B, H * W, self.num_heads, -1)
+
+        # Torch's SDPA expects [B, nheads, H*W, C] so we transpose
+        x = F.scaled_dot_product_attention(
+            q.transpose(1, 2),
+            k.transpose(1, 2),
+            v.transpose(1, 2),
+        )
+        # Transpose back
+        x = x.transpose(1, 2)
+        x = x.reshape(B, H, W, -1)
+
+        x = self.proj(x)
+
+        return x
+
+
+class MultiScaleBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        dim_out: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        drop_path: float = 0.0,
+        norm_layer: Union[nn.Module, str] = "LayerNorm",
+        q_stride: Tuple[int, int] = None,
+        act_layer: nn.Module = nn.GELU,
+        window_size: int = 0,
+    ):
+        super().__init__()
+
+        if isinstance(norm_layer, str):
+            norm_layer = partial(getattr(nn, norm_layer), eps=1e-6)
+
+        self.dim = dim
+        self.dim_out = dim_out
+        self.norm1 = norm_layer(dim)
+
+        self.window_size = window_size
+
+        self.pool, self.q_stride = None, q_stride
+        if self.q_stride:
+            self.pool = nn.MaxPool2d(
+                kernel_size=q_stride, stride=q_stride, ceil_mode=False
+            )
+
+        self.attn = MultiScaleAttention(
+            dim,
+            dim_out,
+            num_heads=num_heads,
+            q_pool=self.pool,
+        )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+        self.norm2 = norm_layer(dim_out)
+        self.mlp = MLP(
+            dim_out,
+            int(dim_out * mlp_ratio),
+            dim_out,
+            num_layers=2,
+            activation=act_layer,
+        )
+
+        if dim != dim_out:
+            self.proj = nn.Linear(dim, dim_out)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        shortcut = x  # B, H, W, C
+        x = self.norm1(x)
+
+        # Skip connection
+        if self.dim != self.dim_out:
+            shortcut = do_pool(self.proj(x), self.pool)
+
+        # Window partition
+        window_size = self.window_size
+        if window_size > 0:
+            H, W = x.shape[1], x.shape[2]
+            x, pad_hw = window_partition(x, window_size)
+
+        # Window Attention + Q Pooling (if stage change)
+        x = self.attn(x)
+        if self.q_stride:
+            # Shapes have changed due to Q pooling
+            window_size = self.window_size // self.q_stride[0]
+            H, W = shortcut.shape[1:3]
+
+            pad_h = (window_size - H % window_size) % window_size
+            pad_w = (window_size - W % window_size) % window_size
+            pad_hw = (H + pad_h, W + pad_w)
+
+        # Reverse window partition
+        if self.window_size > 0:
+            x = window_unpartition(x, window_size, pad_hw, (H, W))
+
+        x = shortcut + self.drop_path(x)
+        # MLP
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+class Hiera(nn.Module):
+    """
+    Reference: https://arxiv.org/abs/2306.00989
+    """
+
+    def __init__(
+        self,
+        embed_dim: int = 96,  # initial embed dim
+        num_heads: int = 1,  # initial number of heads
+        drop_path_rate: float = 0.0,  # stochastic depth
+        q_pool: int = 3,  # number of q_pool stages
+        q_stride: Tuple[int, int] = (2, 2),  # downsample stride bet. stages
+        stages: Tuple[int, ...] = (2, 3, 16, 3),  # blocks per stage
+        dim_mul: float = 2.0,  # dim_mul factor at stage shift
+        head_mul: float = 2.0,  # head_mul factor at stage shift
+        window_pos_embed_bkg_spatial_size: Tuple[int, int] = (14, 14),
+        # window size per stage, when not using global att.
+        window_spec: Tuple[int, ...] = (
+            8,
+            4,
+            14,
+            7,
+        ),
+        # global attn in these blocks
+        global_att_blocks: Tuple[int, ...] = (
+            12,
+            16,
+            20,
+        ),
+        weights_path=None,
+        return_interm_layers=True,  # return feats from every stage
+    ):
+        super().__init__()
+
+        assert len(stages) == len(window_spec)
+        self.window_spec = window_spec
+
+        depth = sum(stages)
+        self.q_stride = q_stride
+        self.stage_ends = [sum(stages[:i]) - 1 for i in range(1, len(stages) + 1)]
+        assert 0 <= q_pool <= len(self.stage_ends[:-1])
+        self.q_pool_blocks = [x + 1 for x in self.stage_ends[:-1]][:q_pool]
+        self.return_interm_layers = return_interm_layers
+
+        self.patch_embed = PatchEmbed(
+            embed_dim=embed_dim,
+        )
+        # Which blocks have global att?
+        self.global_att_blocks = global_att_blocks
+
+        # Windowed positional embedding (https://arxiv.org/abs/2311.05613)
+        self.window_pos_embed_bkg_spatial_size = window_pos_embed_bkg_spatial_size
+        self.pos_embed = nn.Parameter(
+            torch.zeros(1, embed_dim, *self.window_pos_embed_bkg_spatial_size)
+        )
+        self.pos_embed_window = nn.Parameter(
+            torch.zeros(1, embed_dim, self.window_spec[0], self.window_spec[0])
+        )
+
+        dpr = [
+            x.item() for x in torch.linspace(0, drop_path_rate, depth)
+        ]  # stochastic depth decay rule
+
+        cur_stage = 1
+        self.blocks = nn.ModuleList()
+
+        for i in range(depth):
+            dim_out = embed_dim
+            # lags by a block, so first block of
+            # next stage uses an initial window size
+            # of previous stage and final window size of current stage
+            window_size = self.window_spec[cur_stage - 1]
+
+            if self.global_att_blocks is not None:
+                window_size = 0 if i in self.global_att_blocks else window_size
+
+            if i - 1 in self.stage_ends:
+                dim_out = int(embed_dim * dim_mul)
+                num_heads = int(num_heads * head_mul)
+                cur_stage += 1
+
+            block = MultiScaleBlock(
+                dim=embed_dim,
+                dim_out=dim_out,
+                num_heads=num_heads,
+                drop_path=dpr[i],
+                q_stride=self.q_stride if i in self.q_pool_blocks else None,
+                window_size=window_size,
+            )
+
+            embed_dim = dim_out
+            self.blocks.append(block)
+
+        self.channel_list = (
+            [self.blocks[i].dim_out for i in self.stage_ends[::-1]]
+            if return_interm_layers
+            else [self.blocks[-1].dim_out]
+        )
+
+        # if weights_path is not None:
+        #     with g_pathmgr.open(weights_path, "rb") as f:
+        #         chkpt = torch.load(f, map_location="cpu")
+        #     logging.info("loading Hiera", self.load_state_dict(chkpt, strict=False))
+
+    def _get_pos_embed(self, hw: Tuple[int, int]) -> torch.Tensor:
+        h, w = hw
+        window_embed = self.pos_embed_window
+        pos_embed = F.interpolate(self.pos_embed, size=(h, w), mode="bicubic")
+        pos_embed = pos_embed + window_embed.tile(
+            [x // y for x, y in zip(pos_embed.shape, window_embed.shape)]
+        )
+        pos_embed = pos_embed.permute(0, 2, 3, 1)
+        return pos_embed
+
+    def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+        x = self.patch_embed(x)
+        # x: (B, H, W, C)
+
+        # Add pos embed
+        x = x + self._get_pos_embed(x.shape[1:3])
+
+        outputs = []
+        for i, blk in enumerate(self.blocks):
+            x = blk(x)
+            if (i == self.stage_ends[-1]) or (
+                i in self.stage_ends and self.return_interm_layers
+            ):
+                feats = x.permute(0, 3, 1, 2)
+                outputs.append(feats)
+
+        return outputs
+
+    def get_layer_id(self, layer_name):
+        # https://github.com/microsoft/unilm/blob/master/beit/optim_factory.py#L33
+        num_layers = self.get_num_layers()
+
+        if layer_name.find("rel_pos") != -1:
+            return num_layers + 1
+        elif layer_name.find("pos_embed") != -1:
+            return 0
+        elif layer_name.find("patch_embed") != -1:
+            return 0
+        elif layer_name.find("blocks") != -1:
+            return int(layer_name.split("blocks")[1].split(".")[1]) + 1
+        else:
+            return num_layers + 1
+
+    def get_num_layers(self) -> int:
+        return len(self.blocks)

custom_nodes/comfyui-segment-anything-2/sam2/modeling/backbones/image_encoder.py

@@ -0,0 +1,134 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import List, Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class ImageEncoder(nn.Module):
+    def __init__(
+        self,
+        trunk: nn.Module,
+        neck: nn.Module,
+        scalp: int = 0,
+    ):
+        super().__init__()
+        self.trunk = trunk
+        self.neck = neck
+        self.scalp = scalp
+        assert (
+            self.trunk.channel_list == self.neck.backbone_channel_list
+        ), f"Channel dims of trunk and neck do not match. Trunk: {self.trunk.channel_list}, neck: {self.neck.backbone_channel_list}"
+
+    def forward(self, sample: torch.Tensor):
+        # Forward through backbone
+        features, pos = self.neck(self.trunk(sample))
+        if self.scalp > 0:
+            # Discard the lowest resolution features
+            features, pos = features[: -self.scalp], pos[: -self.scalp]
+
+        src = features[-1]
+        output = {
+            "vision_features": src,
+            "vision_pos_enc": pos,
+            "backbone_fpn": features,
+        }
+        return output
+
+
+class FpnNeck(nn.Module):
+    """
+    A modified variant of Feature Pyramid Network (FPN) neck
+    (we remove output conv and also do bicubic interpolation similar to ViT
+    pos embed interpolation)
+    """
+
+    def __init__(
+        self,
+        position_encoding: nn.Module,
+        d_model: int,
+        backbone_channel_list: List[int],
+        kernel_size: int = 1,
+        stride: int = 1,
+        padding: int = 0,
+        fpn_interp_model: str = "bilinear",
+        fuse_type: str = "sum",
+        fpn_top_down_levels: Optional[List[int]] = None,
+    ):
+        """Initialize the neck
+        :param trunk: the backbone
+        :param position_encoding: the positional encoding to use
+        :param d_model: the dimension of the model
+        :param neck_norm: the normalization to use
+        """
+        super().__init__()
+        self.position_encoding = position_encoding
+        self.convs = nn.ModuleList()
+        self.backbone_channel_list = backbone_channel_list
+        self.d_model = d_model
+        for dim in backbone_channel_list:
+            current = nn.Sequential()
+            current.add_module(
+                "conv",
+                nn.Conv2d(
+                    in_channels=dim,
+                    out_channels=d_model,
+                    kernel_size=kernel_size,
+                    stride=stride,
+                    padding=padding,
+                ),
+            )
+
+            self.convs.append(current)
+        self.fpn_interp_model = fpn_interp_model
+        assert fuse_type in ["sum", "avg"]
+        self.fuse_type = fuse_type
+
+        # levels to have top-down features in its outputs
+        # e.g. if fpn_top_down_levels is [2, 3], then only outputs of level 2 and 3
+        # have top-down propagation, while outputs of level 0 and level 1 have only
+        # lateral features from the same backbone level.
+        if fpn_top_down_levels is None:
+            # default is to have top-down features on all levels
+            fpn_top_down_levels = range(len(self.convs))
+        self.fpn_top_down_levels = list(fpn_top_down_levels)
+
+    def forward(self, xs: List[torch.Tensor]):
+
+        out = [None] * len(self.convs)
+        pos = [None] * len(self.convs)
+        assert len(xs) == len(self.convs)
+        # fpn forward pass
+        # see https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/fpn.py
+        prev_features = None
+        # forward in top-down order (from low to high resolution)
+        n = len(self.convs) - 1
+        for i in range(n, -1, -1):
+            x = xs[i]
+            lateral_features = self.convs[n - i](x)
+            if i in self.fpn_top_down_levels and prev_features is not None:
+                top_down_features = F.interpolate(
+                    prev_features.to(dtype=torch.float32),
+                    scale_factor=2.0,
+                    mode=self.fpn_interp_model,
+                    align_corners=(
+                        None if self.fpn_interp_model == "nearest" else False
+                    ),
+                    antialias=False,
+                )
+                prev_features = lateral_features + top_down_features
+                if self.fuse_type == "avg":
+                    prev_features /= 2
+            else:
+                prev_features = lateral_features
+            x_out = prev_features
+            out[i] = x_out
+            pos[i] = self.position_encoding(x_out).to(x_out.dtype)
+
+        return out, pos

custom_nodes/comfyui-segment-anything-2/sam2/modeling/backbones/utils.py

@@ -0,0 +1,95 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Some utilities for backbones, in particular for windowing"""
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def window_partition(x, window_size):
+    """
+    Partition into non-overlapping windows with padding if needed.
+    Args:
+        x (tensor): input tokens with [B, H, W, C].
+        window_size (int): window size.
+    Returns:
+        windows: windows after partition with [B * num_windows, window_size, window_size, C].
+        (Hp, Wp): padded height and width before partition
+    """
+    B, H, W, C = x.shape
+
+    pad_h = (window_size - H % window_size) % window_size
+    pad_w = (window_size - W % window_size) % window_size
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
+    Hp, Wp = H + pad_h, W + pad_w
+
+    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
+    windows = (
+        x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    )
+    return windows, (Hp, Wp)
+
+
+def window_unpartition(windows, window_size, pad_hw, hw):
+    """
+    Window unpartition into original sequences and removing padding.
+    Args:
+        x (tensor): input tokens with [B * num_windows, window_size, window_size, C].
+        window_size (int): window size.
+        pad_hw (Tuple): padded height and width (Hp, Wp).
+        hw (Tuple): original height and width (H, W) before padding.
+    Returns:
+        x: unpartitioned sequences with [B, H, W, C].
+    """
+    Hp, Wp = pad_hw
+    H, W = hw
+    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
+    x = windows.view(
+        B, Hp // window_size, Wp // window_size, window_size, window_size, -1
+    )
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
+
+    if Hp > H or Wp > W:
+        x = x[:, :H, :W, :].contiguous()
+    return x
+
+
+class PatchEmbed(nn.Module):
+    """
+    Image to Patch Embedding.
+    """
+
+    def __init__(
+        self,
+        kernel_size: Tuple[int, ...] = (7, 7),
+        stride: Tuple[int, ...] = (4, 4),
+        padding: Tuple[int, ...] = (3, 3),
+        in_chans: int = 3,
+        embed_dim: int = 768,
+    ):
+        """
+        Args:
+            kernel_size (Tuple): kernel size of the projection layer.
+            stride (Tuple): stride of the projection layer.
+            padding (Tuple): padding size of the projection layer.
+            in_chans (int): Number of input image channels.
+            embed_dim (int):  embed_dim (int): Patch embedding dimension.
+        """
+        super().__init__()
+        self.proj = nn.Conv2d(
+            in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        # B C H W -> B H W C
+        x = x.permute(0, 2, 3, 1)
+        return x

custom_nodes/comfyui-segment-anything-2/sam2/modeling/memory_attention.py

@@ -0,0 +1,169 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Optional
+
+import torch
+from torch import nn, Tensor
+
+from ...sam2.modeling.sam.transformer import RoPEAttention
+
+from ...sam2.modeling.sam2_utils import get_activation_fn, get_clones
+
+
+class MemoryAttentionLayer(nn.Module):
+
+    def __init__(
+        self,
+        activation: str,
+        cross_attention: nn.Module,
+        d_model: int,
+        dim_feedforward: int,
+        dropout: float,
+        pos_enc_at_attn: bool,
+        pos_enc_at_cross_attn_keys: bool,
+        pos_enc_at_cross_attn_queries: bool,
+        self_attention: nn.Module,
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.dim_feedforward = dim_feedforward
+        self.dropout_value = dropout
+        self.self_attn = self_attention
+        self.cross_attn_image = cross_attention
+
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation_str = activation
+        self.activation = get_activation_fn(activation)
+
+        # Where to add pos enc
+        self.pos_enc_at_attn = pos_enc_at_attn
+        self.pos_enc_at_cross_attn_queries = pos_enc_at_cross_attn_queries
+        self.pos_enc_at_cross_attn_keys = pos_enc_at_cross_attn_keys
+
+    def _forward_sa(self, tgt, query_pos):
+        # Self-Attention
+        tgt2 = self.norm1(tgt)
+        q = k = tgt2 + query_pos if self.pos_enc_at_attn else tgt2
+        tgt2 = self.self_attn(q, k, v=tgt2)
+        tgt = tgt + self.dropout1(tgt2)
+        return tgt
+
+    def _forward_ca(self, tgt, memory, query_pos, pos, num_k_exclude_rope=0):
+        kwds = {}
+        if num_k_exclude_rope > 0:
+            assert isinstance(self.cross_attn_image, RoPEAttention)
+            kwds = {"num_k_exclude_rope": num_k_exclude_rope}
+
+        # Cross-Attention
+        tgt2 = self.norm2(tgt)
+        tgt2 = self.cross_attn_image(
+            q=tgt2 + query_pos if self.pos_enc_at_cross_attn_queries else tgt2,
+            k=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
+            v=memory,
+            **kwds,
+        )
+        tgt = tgt + self.dropout2(tgt2)
+        return tgt
+
+    def forward(
+        self,
+        tgt,
+        memory,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        num_k_exclude_rope: int = 0,
+    ) -> torch.Tensor:
+
+        # Self-Attn, Cross-Attn
+        tgt = self._forward_sa(tgt, query_pos)
+        tgt = self._forward_ca(tgt, memory, query_pos, pos, num_k_exclude_rope)
+        # MLP
+        tgt2 = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
+        tgt = tgt + self.dropout3(tgt2)
+        return tgt
+
+
+class MemoryAttention(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        pos_enc_at_input: bool,
+        layer: nn.Module,
+        num_layers: int,
+        batch_first: bool = True,  # Do layers expect batch first input?
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.layers = get_clones(layer, num_layers)
+        self.num_layers = num_layers
+        self.norm = nn.LayerNorm(d_model)
+        self.pos_enc_at_input = pos_enc_at_input
+        self.batch_first = batch_first
+
+    def forward(
+        self,
+        curr: torch.Tensor,  # self-attention inputs
+        memory: torch.Tensor,  # cross-attention inputs
+        curr_pos: Optional[Tensor] = None,  # pos_enc for self-attention inputs
+        memory_pos: Optional[Tensor] = None,  # pos_enc for cross-attention inputs
+        num_obj_ptr_tokens: int = 0,  # number of object pointer *tokens*
+    ):
+        if isinstance(curr, list):
+            assert isinstance(curr_pos, list)
+            assert len(curr) == len(curr_pos) == 1
+            curr, curr_pos = (
+                curr[0],
+                curr_pos[0],
+            )
+
+        assert (
+            curr.shape[1] == memory.shape[1]
+        ), "Batch size must be the same for curr and memory"
+
+        output = curr
+        if self.pos_enc_at_input and curr_pos is not None:
+            output = output + 0.1 * curr_pos
+
+        if self.batch_first:
+            # Convert to batch first
+            output = output.transpose(0, 1)
+            curr_pos = curr_pos.transpose(0, 1)
+            memory = memory.transpose(0, 1)
+            memory_pos = memory_pos.transpose(0, 1)
+
+        for layer in self.layers:
+            kwds = {}
+            if isinstance(layer.cross_attn_image, RoPEAttention):
+                kwds = {"num_k_exclude_rope": num_obj_ptr_tokens}
+
+            output = layer(
+                tgt=output,
+                memory=memory,
+                pos=memory_pos,
+                query_pos=curr_pos,
+                **kwds,
+            )
+        normed_output = self.norm(output)
+
+        if self.batch_first:
+            # Convert back to seq first
+            normed_output = normed_output.transpose(0, 1)
+            curr_pos = curr_pos.transpose(0, 1)
+
+        return normed_output

custom_nodes/comfyui-segment-anything-2/sam2/modeling/memory_encoder.py

@@ -0,0 +1,181 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ...sam2.modeling.sam2_utils import DropPath, get_clones, LayerNorm2d
+
+
+class MaskDownSampler(nn.Module):
+    """
+    Progressively downsample a mask by total_stride, each time by stride.
+    Note that LayerNorm is applied per *token*, like in ViT.
+
+    With each downsample (by a factor stride**2), channel capacity increases by the same factor.
+    In the end, we linearly project to embed_dim channels.
+    """
+
+    def __init__(
+        self,
+        embed_dim=256,
+        kernel_size=4,
+        stride=4,
+        padding=0,
+        total_stride=16,
+        activation=nn.GELU,
+    ):
+        super().__init__()
+        num_layers = int(math.log2(total_stride) // math.log2(stride))
+        assert stride**num_layers == total_stride
+        self.encoder = nn.Sequential()
+        mask_in_chans, mask_out_chans = 1, 1
+        for _ in range(num_layers):
+            mask_out_chans = mask_in_chans * (stride**2)
+            self.encoder.append(
+                nn.Conv2d(
+                    mask_in_chans,
+                    mask_out_chans,
+                    kernel_size=kernel_size,
+                    stride=stride,
+                    padding=padding,
+                )
+            )
+            self.encoder.append(LayerNorm2d(mask_out_chans))
+            self.encoder.append(activation())
+            mask_in_chans = mask_out_chans
+
+        self.encoder.append(nn.Conv2d(mask_out_chans, embed_dim, kernel_size=1))
+
+    def forward(self, x):
+        return self.encoder(x)
+
+
+# Lightly adapted from ConvNext (https://github.com/facebookresearch/ConvNeXt)
+class CXBlock(nn.Module):
+    r"""ConvNeXt Block. There are two equivalent implementations:
+    (1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)
+    (2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back
+    We use (2) as we find it slightly faster in PyTorch
+
+    Args:
+        dim (int): Number of input channels.
+        drop_path (float): Stochastic depth rate. Default: 0.0
+        layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
+    """
+
+    def __init__(
+        self,
+        dim,
+        kernel_size=7,
+        padding=3,
+        drop_path=0.0,
+        layer_scale_init_value=1e-6,
+        use_dwconv=True,
+    ):
+        super().__init__()
+        self.dwconv = nn.Conv2d(
+            dim,
+            dim,
+            kernel_size=kernel_size,
+            padding=padding,
+            groups=dim if use_dwconv else 1,
+        )  # depthwise conv
+        self.norm = LayerNorm2d(dim, eps=1e-6)
+        self.pwconv1 = nn.Linear(
+            dim, 4 * dim
+        )  # pointwise/1x1 convs, implemented with linear layers
+        self.act = nn.GELU()
+        self.pwconv2 = nn.Linear(4 * dim, dim)
+        self.gamma = (
+            nn.Parameter(layer_scale_init_value * torch.ones((dim)), requires_grad=True)
+            if layer_scale_init_value > 0
+            else None
+        )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+    def forward(self, x):
+        input = x
+        x = self.dwconv(x)
+        x = self.norm(x)
+        x = x.permute(0, 2, 3, 1)  # (N, C, H, W) -> (N, H, W, C)
+        x = self.pwconv1(x)
+        x = self.act(x)
+        x = self.pwconv2(x)
+        if self.gamma is not None:
+            x = self.gamma * x
+        x = x.permute(0, 3, 1, 2)  # (N, H, W, C) -> (N, C, H, W)
+
+        x = input + self.drop_path(x)
+        return x
+
+
+class Fuser(nn.Module):
+    def __init__(self, layer, num_layers, dim=None, input_projection=False):
+        super().__init__()
+        self.proj = nn.Identity()
+        self.layers = get_clones(layer, num_layers)
+
+        if input_projection:
+            assert dim is not None
+            self.proj = nn.Conv2d(dim, dim, kernel_size=1)
+
+    def forward(self, x):
+        # normally x: (N, C, H, W)
+        x = self.proj(x)
+        for layer in self.layers:
+            x = layer(x)
+        return x
+
+
+class MemoryEncoder(nn.Module):
+    def __init__(
+        self,
+        out_dim,
+        mask_downsampler,
+        fuser,
+        position_encoding,
+        in_dim=256,  # in_dim of pix_feats
+    ):
+        super().__init__()
+
+        self.mask_downsampler = mask_downsampler
+
+        self.pix_feat_proj = nn.Conv2d(in_dim, in_dim, kernel_size=1)
+        self.fuser = fuser
+        self.position_encoding = position_encoding
+        self.out_proj = nn.Identity()
+        if out_dim != in_dim:
+            self.out_proj = nn.Conv2d(in_dim, out_dim, kernel_size=1)
+
+    def forward(
+        self,
+        pix_feat: torch.Tensor,
+        masks: torch.Tensor,
+        skip_mask_sigmoid: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        ## Process masks
+        # sigmoid, so that less domain shift from gt masks which are bool
+        if not skip_mask_sigmoid:
+            masks = F.sigmoid(masks)
+        masks = self.mask_downsampler(masks)
+
+        ## Fuse pix_feats and downsampled masks
+        # in case the visual features are on CPU, cast them to CUDA
+        pix_feat = pix_feat.to(masks.device)
+
+        x = self.pix_feat_proj(pix_feat)
+        x = x + masks
+        x = self.fuser(x)
+        x = self.out_proj(x)
+
+        pos = self.position_encoding(x).to(x.dtype)
+
+        return {"vision_features": x, "vision_pos_enc": [pos]}

custom_nodes/comfyui-segment-anything-2/sam2/modeling/position_encoding.py

@@ -0,0 +1,220 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from typing import Any, Optional, Tuple
+
+import numpy as np
+
+import torch
+from torch import nn
+
+
+class PositionEmbeddingSine(nn.Module):
+    """
+    This is a more standard version of the position embedding, very similar to the one
+    used by the Attention is all you need paper, generalized to work on images.
+    """
+
+    def __init__(
+        self,
+        num_pos_feats,
+        temperature: int = 10000,
+        normalize: bool = True,
+        scale: Optional[float] = None,
+    ):
+        super().__init__()
+        assert num_pos_feats % 2 == 0, "Expecting even model width"
+        self.num_pos_feats = num_pos_feats // 2
+        self.temperature = temperature
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+
+        self.cache = {}
+
+    def _encode_xy(self, x, y):
+        # The positions are expected to be normalized
+        assert len(x) == len(y) and x.ndim == y.ndim == 1
+        x_embed = x * self.scale
+        y_embed = y * self.scale
+
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+
+        pos_x = x_embed[:, None] / dim_t
+        pos_y = y_embed[:, None] / dim_t
+        pos_x = torch.stack(
+            (pos_x[:, 0::2].sin(), pos_x[:, 1::2].cos()), dim=2
+        ).flatten(1)
+        pos_y = torch.stack(
+            (pos_y[:, 0::2].sin(), pos_y[:, 1::2].cos()), dim=2
+        ).flatten(1)
+        return pos_x, pos_y
+
+    @torch.no_grad()
+    def encode_boxes(self, x, y, w, h):
+        pos_x, pos_y = self._encode_xy(x, y)
+        pos = torch.cat((pos_y, pos_x, h[:, None], w[:, None]), dim=1)
+        return pos
+
+    encode = encode_boxes  # Backwards compatibility
+
+    @torch.no_grad()
+    def encode_points(self, x, y, labels):
+        (bx, nx), (by, ny), (bl, nl) = x.shape, y.shape, labels.shape
+        assert bx == by and nx == ny and bx == bl and nx == nl
+        pos_x, pos_y = self._encode_xy(x.flatten(), y.flatten())
+        pos_x, pos_y = pos_x.reshape(bx, nx, -1), pos_y.reshape(by, ny, -1)
+        pos = torch.cat((pos_y, pos_x, labels[:, :, None]), dim=2)
+        return pos
+
+    @torch.no_grad()
+    def forward(self, x: torch.Tensor):
+        cache_key = (x.shape[-2], x.shape[-1])
+        if cache_key in self.cache:
+            return self.cache[cache_key][None].repeat(x.shape[0], 1, 1, 1)
+        y_embed = (
+            torch.arange(1, x.shape[-2] + 1, dtype=torch.float32, device=x.device)
+            .view(1, -1, 1)
+            .repeat(x.shape[0], 1, x.shape[-1])
+        )
+        x_embed = (
+            torch.arange(1, x.shape[-1] + 1, dtype=torch.float32, device=x.device)
+            .view(1, 1, -1)
+            .repeat(x.shape[0], x.shape[-2], 1)
+        )
+
+        if self.normalize:
+            eps = 1e-6
+            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+        pos_x = torch.stack(
+            (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos_y = torch.stack(
+            (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+        self.cache[cache_key] = pos[0]
+        return pos
+
+
+class PositionEmbeddingRandom(nn.Module):
+    """
+    Positional encoding using random spatial frequencies.
+    """
+
+    def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:
+        super().__init__()
+        if scale is None or scale <= 0.0:
+            scale = 1.0
+        self.register_buffer(
+            "positional_encoding_gaussian_matrix",
+            scale * torch.randn((2, num_pos_feats)),
+        )
+
+    def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
+        """Positionally encode points that are normalized to [0,1]."""
+        # assuming coords are in [0, 1]^2 square and have d_1 x ... x d_n x 2 shape
+        coords = 2 * coords - 1
+        coords = coords @ self.positional_encoding_gaussian_matrix
+        coords = 2 * np.pi * coords
+        # outputs d_1 x ... x d_n x C shape
+        return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)
+
+    def forward(self, size: Tuple[int, int]) -> torch.Tensor:
+        """Generate positional encoding for a grid of the specified size."""
+        h, w = size
+        device: Any = self.positional_encoding_gaussian_matrix.device
+        grid = torch.ones((h, w), device=device, dtype=torch.float32)
+        y_embed = grid.cumsum(dim=0) - 0.5
+        x_embed = grid.cumsum(dim=1) - 0.5
+        y_embed = y_embed / h
+        x_embed = x_embed / w
+
+        pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))
+        return pe.permute(2, 0, 1)  # C x H x W
+
+    def forward_with_coords(
+        self, coords_input: torch.Tensor, image_size: Tuple[int, int]
+    ) -> torch.Tensor:
+        """Positionally encode points that are not normalized to [0,1]."""
+        coords = coords_input.clone()
+        coords[:, :, 0] = coords[:, :, 0] / image_size[1]
+        coords[:, :, 1] = coords[:, :, 1] / image_size[0]
+        return self._pe_encoding(coords.to(torch.float))  # B x N x C
+
+
+# Rotary Positional Encoding, adapted from:
+# 1. https://github.com/meta-llama/codellama/blob/main/llama/model.py
+# 2. https://github.com/naver-ai/rope-vit
+# 3. https://github.com/lucidrains/rotary-embedding-torch
+
+
+def init_t_xy(end_x: int, end_y: int):
+    t = torch.arange(end_x * end_y, dtype=torch.float32)
+    t_x = (t % end_x).float()
+    t_y = torch.div(t, end_x, rounding_mode="floor").float()
+    return t_x, t_y
+
+
+def compute_axial_cis(dim: int, end_x: int, end_y: int, theta: float = 10000.0):
+    freqs_x = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
+    freqs_y = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
+
+    t_x, t_y = init_t_xy(end_x, end_y)
+    freqs_x = torch.outer(t_x, freqs_x)
+    freqs_y = torch.outer(t_y, freqs_y)
+    freqs_cis_x = torch.polar(torch.ones_like(freqs_x), freqs_x)
+    freqs_cis_y = torch.polar(torch.ones_like(freqs_y), freqs_y)
+    return torch.cat([freqs_cis_x, freqs_cis_y], dim=-1)
+
+
+def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor):
+    ndim = x.ndim
+    assert 0 <= 1 < ndim
+    assert freqs_cis.shape == (x.shape[-2], x.shape[-1])
+    shape = [d if i >= ndim - 2 else 1 for i, d in enumerate(x.shape)]
+    return freqs_cis.view(*shape)
+
+
+def apply_rotary_enc(
+    xq: torch.Tensor,
+    xk: torch.Tensor,
+    freqs_cis: torch.Tensor,
+    repeat_freqs_k: bool = False,
+):
+    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
+    xk_ = (
+        torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
+        if xk.shape[-2] != 0
+        else None
+    )
+    freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
+    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
+    if xk_ is None:
+        # no keys to rotate, due to dropout
+        return xq_out.type_as(xq).to(xq.device), xk
+    # repeat freqs along seq_len dim to match k seq_len
+    if repeat_freqs_k:
+        r = xk_.shape[-2] // xq_.shape[-2]
+        if freqs_cis.is_complex() and freqs_cis.device.type == "mps":
+            # MPS doesn't support repeat on complex; cat works fine.
+            freqs_cis = torch.cat([freqs_cis] * r, dim=-2)
+        else:
+            freqs_cis = freqs_cis.repeat(*([1] * (freqs_cis.ndim - 2)), r, 1)
+    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
+    return xq_out.type_as(xq).to(xq.device), xk_out.type_as(xk).to(xk.device)

custom_nodes/comfyui-segment-anything-2/sam2/modeling/sam/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

custom_nodes/comfyui-segment-anything-2/sam2/modeling/sam/mask_decoder.py

@@ -0,0 +1,295 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import List, Optional, Tuple, Type
+
+import torch
+from torch import nn
+
+from ....sam2.modeling.sam2_utils import LayerNorm2d, MLP
+
+
+class MaskDecoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        transformer_dim: int,
+        transformer: nn.Module,
+        num_multimask_outputs: int = 3,
+        activation: Type[nn.Module] = nn.GELU,
+        iou_head_depth: int = 3,
+        iou_head_hidden_dim: int = 256,
+        use_high_res_features: bool = False,
+        iou_prediction_use_sigmoid=False,
+        dynamic_multimask_via_stability=False,
+        dynamic_multimask_stability_delta=0.05,
+        dynamic_multimask_stability_thresh=0.98,
+        pred_obj_scores: bool = False,
+        pred_obj_scores_mlp: bool = False,
+        use_multimask_token_for_obj_ptr: bool = False,
+    ) -> None:
+        """
+        Predicts masks given an image and prompt embeddings, using a
+        transformer architecture.
+
+        Arguments:
+          transformer_dim (int): the channel dimension of the transformer
+          transformer (nn.Module): the transformer used to predict masks
+          num_multimask_outputs (int): the number of masks to predict
+            when disambiguating masks
+          activation (nn.Module): the type of activation to use when
+            upscaling masks
+          iou_head_depth (int): the depth of the MLP used to predict
+            mask quality
+          iou_head_hidden_dim (int): the hidden dimension of the MLP
+            used to predict mask quality
+        """
+        super().__init__()
+        self.transformer_dim = transformer_dim
+        self.transformer = transformer
+
+        self.num_multimask_outputs = num_multimask_outputs
+
+        self.iou_token = nn.Embedding(1, transformer_dim)
+        self.num_mask_tokens = num_multimask_outputs + 1
+        self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)
+
+        self.pred_obj_scores = pred_obj_scores
+        if self.pred_obj_scores:
+            self.obj_score_token = nn.Embedding(1, transformer_dim)
+        self.use_multimask_token_for_obj_ptr = use_multimask_token_for_obj_ptr
+
+        self.output_upscaling = nn.Sequential(
+            nn.ConvTranspose2d(
+                transformer_dim, transformer_dim // 4, kernel_size=2, stride=2
+            ),
+            LayerNorm2d(transformer_dim // 4),
+            activation(),
+            nn.ConvTranspose2d(
+                transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2
+            ),
+            activation(),
+        )
+        self.use_high_res_features = use_high_res_features
+        if use_high_res_features:
+            self.conv_s0 = nn.Conv2d(
+                transformer_dim, transformer_dim // 8, kernel_size=1, stride=1
+            )
+            self.conv_s1 = nn.Conv2d(
+                transformer_dim, transformer_dim // 4, kernel_size=1, stride=1
+            )
+
+        self.output_hypernetworks_mlps = nn.ModuleList(
+            [
+                MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3)
+                for i in range(self.num_mask_tokens)
+            ]
+        )
+
+        self.iou_prediction_head = MLP(
+            transformer_dim,
+            iou_head_hidden_dim,
+            self.num_mask_tokens,
+            iou_head_depth,
+            sigmoid_output=iou_prediction_use_sigmoid,
+        )
+        if self.pred_obj_scores:
+            self.pred_obj_score_head = nn.Linear(transformer_dim, 1)
+            if pred_obj_scores_mlp:
+                self.pred_obj_score_head = MLP(transformer_dim, transformer_dim, 1, 3)
+
+        # When outputting a single mask, optionally we can dynamically fall back to the best
+        # multimask output token if the single mask output token gives low stability scores.
+        self.dynamic_multimask_via_stability = dynamic_multimask_via_stability
+        self.dynamic_multimask_stability_delta = dynamic_multimask_stability_delta
+        self.dynamic_multimask_stability_thresh = dynamic_multimask_stability_thresh
+
+    def forward(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+        multimask_output: bool,
+        repeat_image: bool,
+        high_res_features: Optional[List[torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Predict masks given image and prompt embeddings.
+
+        Arguments:
+          image_embeddings (torch.Tensor): the embeddings from the image encoder
+          image_pe (torch.Tensor): positional encoding with the shape of image_embeddings
+          sparse_prompt_embeddings (torch.Tensor): the embeddings of the points and boxes
+          dense_prompt_embeddings (torch.Tensor): the embeddings of the mask inputs
+          multimask_output (bool): Whether to return multiple masks or a single
+            mask.
+
+        Returns:
+          torch.Tensor: batched predicted masks
+          torch.Tensor: batched predictions of mask quality
+          torch.Tensor: batched SAM token for mask output
+        """
+        masks, iou_pred, mask_tokens_out, object_score_logits = self.predict_masks(
+            image_embeddings=image_embeddings,
+            image_pe=image_pe,
+            sparse_prompt_embeddings=sparse_prompt_embeddings,
+            dense_prompt_embeddings=dense_prompt_embeddings,
+            repeat_image=repeat_image,
+            high_res_features=high_res_features,
+        )
+
+        # Select the correct mask or masks for output
+        if multimask_output:
+            masks = masks[:, 1:, :, :]
+            iou_pred = iou_pred[:, 1:]
+        elif self.dynamic_multimask_via_stability and not self.training:
+            masks, iou_pred = self._dynamic_multimask_via_stability(masks, iou_pred)
+        else:
+            masks = masks[:, 0:1, :, :]
+            iou_pred = iou_pred[:, 0:1]
+
+        if multimask_output and self.use_multimask_token_for_obj_ptr:
+            sam_tokens_out = mask_tokens_out[:, 1:]  # [b, 3, c] shape
+        else:
+            # Take the mask output token. Here we *always* use the token for single mask output.
+            # At test time, even if we track after 1-click (and using multimask_output=True),
+            # we still take the single mask token here. The rationale is that we always track
+            # after multiple clicks during training, so the past tokens seen during training
+            # are always the single mask token (and we'll let it be the object-memory token).
+            sam_tokens_out = mask_tokens_out[:, 0:1]  # [b, 1, c] shape
+
+        # Prepare output
+        return masks, iou_pred, sam_tokens_out, object_score_logits
+
+    def predict_masks(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+        repeat_image: bool,
+        high_res_features: Optional[List[torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Predicts masks. See 'forward' for more details."""
+        # Concatenate output tokens
+        s = 0
+        if self.pred_obj_scores:
+            output_tokens = torch.cat(
+                [
+                    self.obj_score_token.weight,
+                    self.iou_token.weight,
+                    self.mask_tokens.weight,
+                ],
+                dim=0,
+            )
+            s = 1
+        else:
+            output_tokens = torch.cat(
+                [self.iou_token.weight, self.mask_tokens.weight], dim=0
+            )
+        output_tokens = output_tokens.unsqueeze(0).expand(
+            sparse_prompt_embeddings.size(0), -1, -1
+        )
+        tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)
+
+        # Expand per-image data in batch direction to be per-mask
+        if repeat_image:
+            src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
+        else:
+            assert image_embeddings.shape[0] == tokens.shape[0]
+            src = image_embeddings
+        src = src + dense_prompt_embeddings
+        assert (
+            image_pe.size(0) == 1
+        ), "image_pe should have size 1 in batch dim (from `get_dense_pe()`)"
+        pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
+        b, c, h, w = src.shape
+
+        # Run the transformer
+        hs, src = self.transformer(src, pos_src, tokens)
+        iou_token_out = hs[:, s, :]
+        mask_tokens_out = hs[:, s + 1 : (s + 1 + self.num_mask_tokens), :]
+
+        # Upscale mask embeddings and predict masks using the mask tokens
+        src = src.transpose(1, 2).view(b, c, h, w)
+        if not self.use_high_res_features:
+            upscaled_embedding = self.output_upscaling(src)
+        else:
+            dc1, ln1, act1, dc2, act2 = self.output_upscaling
+            feat_s0, feat_s1 = high_res_features
+            upscaled_embedding = act1(ln1(dc1(src) + feat_s1))
+            upscaled_embedding = act2(dc2(upscaled_embedding) + feat_s0)
+
+        hyper_in_list: List[torch.Tensor] = []
+        for i in range(self.num_mask_tokens):
+            hyper_in_list.append(
+                self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :])
+            )
+        hyper_in = torch.stack(hyper_in_list, dim=1)
+        b, c, h, w = upscaled_embedding.shape
+        masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)
+
+        # Generate mask quality predictions
+        iou_pred = self.iou_prediction_head(iou_token_out)
+        if self.pred_obj_scores:
+            assert s == 1
+            object_score_logits = self.pred_obj_score_head(hs[:, 0, :])
+        else:
+            # Obj scores logits - default to 10.0, i.e. assuming the object is present, sigmoid(10)=1
+            object_score_logits = 10.0 * iou_pred.new_ones(iou_pred.shape[0], 1)
+
+        return masks, iou_pred, mask_tokens_out, object_score_logits
+
+    def _get_stability_scores(self, mask_logits):
+        """
+        Compute stability scores of the mask logits based on the IoU between upper and
+        lower thresholds.
+        """
+        mask_logits = mask_logits.flatten(-2)
+        stability_delta = self.dynamic_multimask_stability_delta
+        area_i = torch.sum(mask_logits > stability_delta, dim=-1).float()
+        area_u = torch.sum(mask_logits > -stability_delta, dim=-1).float()
+        stability_scores = torch.where(area_u > 0, area_i / area_u, 1.0)
+        return stability_scores
+
+    def _dynamic_multimask_via_stability(self, all_mask_logits, all_iou_scores):
+        """
+        When outputting a single mask, if the stability score from the current single-mask
+        output (based on output token 0) falls below a threshold, we instead select from
+        multi-mask outputs (based on output token 1~3) the mask with the highest predicted
+        IoU score. This is intended to ensure a valid mask for both clicking and tracking.
+        """
+        # The best mask from multimask output tokens (1~3)
+        multimask_logits = all_mask_logits[:, 1:, :, :]
+        multimask_iou_scores = all_iou_scores[:, 1:]
+        best_scores_inds = torch.argmax(multimask_iou_scores, dim=-1)
+        batch_inds = torch.arange(
+            multimask_iou_scores.size(0), device=all_iou_scores.device
+        )
+        best_multimask_logits = multimask_logits[batch_inds, best_scores_inds]
+        best_multimask_logits = best_multimask_logits.unsqueeze(1)
+        best_multimask_iou_scores = multimask_iou_scores[batch_inds, best_scores_inds]
+        best_multimask_iou_scores = best_multimask_iou_scores.unsqueeze(1)
+
+        # The mask from singlemask output token 0 and its stability score
+        singlemask_logits = all_mask_logits[:, 0:1, :, :]
+        singlemask_iou_scores = all_iou_scores[:, 0:1]
+        stability_scores = self._get_stability_scores(singlemask_logits)
+        is_stable = stability_scores >= self.dynamic_multimask_stability_thresh
+
+        # Dynamically fall back to best multimask output upon low stability scores.
+        mask_logits_out = torch.where(
+            is_stable[..., None, None].expand_as(singlemask_logits),
+            singlemask_logits,
+            best_multimask_logits,
+        )
+        iou_scores_out = torch.where(
+            is_stable.expand_as(singlemask_iou_scores),
+            singlemask_iou_scores,
+            best_multimask_iou_scores,
+        )
+        return mask_logits_out, iou_scores_out

custom_nodes/comfyui-segment-anything-2/sam2/modeling/sam/prompt_encoder.py

@@ -0,0 +1,182 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Optional, Tuple, Type
+
+import torch
+from torch import nn
+
+from ....sam2.modeling.position_encoding import PositionEmbeddingRandom
+
+from ....sam2.modeling.sam2_utils import LayerNorm2d
+
+
+class PromptEncoder(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        image_embedding_size: Tuple[int, int],
+        input_image_size: Tuple[int, int],
+        mask_in_chans: int,
+        activation: Type[nn.Module] = nn.GELU,
+    ) -> None:
+        """
+        Encodes prompts for input to SAM's mask decoder.
+
+        Arguments:
+          embed_dim (int): The prompts' embedding dimension
+          image_embedding_size (tuple(int, int)): The spatial size of the
+            image embedding, as (H, W).
+          input_image_size (int): The padded size of the image as input
+            to the image encoder, as (H, W).
+          mask_in_chans (int): The number of hidden channels used for
+            encoding input masks.
+          activation (nn.Module): The activation to use when encoding
+            input masks.
+        """
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.input_image_size = input_image_size
+        self.image_embedding_size = image_embedding_size
+        self.pe_layer = PositionEmbeddingRandom(embed_dim // 2)
+
+        self.num_point_embeddings: int = 4  # pos/neg point + 2 box corners
+        point_embeddings = [
+            nn.Embedding(1, embed_dim) for i in range(self.num_point_embeddings)
+        ]
+        self.point_embeddings = nn.ModuleList(point_embeddings)
+        self.not_a_point_embed = nn.Embedding(1, embed_dim)
+
+        self.mask_input_size = (
+            4 * image_embedding_size[0],
+            4 * image_embedding_size[1],
+        )
+        self.mask_downscaling = nn.Sequential(
+            nn.Conv2d(1, mask_in_chans // 4, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans // 4),
+            activation(),
+            nn.Conv2d(mask_in_chans // 4, mask_in_chans, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans),
+            activation(),
+            nn.Conv2d(mask_in_chans, embed_dim, kernel_size=1),
+        )
+        self.no_mask_embed = nn.Embedding(1, embed_dim)
+
+    def get_dense_pe(self) -> torch.Tensor:
+        """
+        Returns the positional encoding used to encode point prompts,
+        applied to a dense set of points the shape of the image encoding.
+
+        Returns:
+          torch.Tensor: Positional encoding with shape
+            1x(embed_dim)x(embedding_h)x(embedding_w)
+        """
+        return self.pe_layer(self.image_embedding_size).unsqueeze(0)
+
+    def _embed_points(
+        self,
+        points: torch.Tensor,
+        labels: torch.Tensor,
+        pad: bool,
+    ) -> torch.Tensor:
+        """Embeds point prompts."""
+        points = points + 0.5  # Shift to center of pixel
+        if pad:
+            padding_point = torch.zeros((points.shape[0], 1, 2), device=points.device)
+            padding_label = -torch.ones((labels.shape[0], 1), device=labels.device)
+            points = torch.cat([points, padding_point], dim=1)
+            labels = torch.cat([labels, padding_label], dim=1)
+        point_embedding = self.pe_layer.forward_with_coords(
+            points, self.input_image_size
+        )
+        point_embedding[labels == -1] = 0.0
+        point_embedding[labels == -1] += self.not_a_point_embed.weight
+        point_embedding[labels == 0] += self.point_embeddings[0].weight
+        point_embedding[labels == 1] += self.point_embeddings[1].weight
+        point_embedding[labels == 2] += self.point_embeddings[2].weight
+        point_embedding[labels == 3] += self.point_embeddings[3].weight
+        return point_embedding
+
+    def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
+        """Embeds box prompts."""
+        boxes = boxes + 0.5  # Shift to center of pixel
+        coords = boxes.reshape(-1, 2, 2)
+        corner_embedding = self.pe_layer.forward_with_coords(
+            coords, self.input_image_size
+        )
+        corner_embedding[:, 0, :] += self.point_embeddings[2].weight
+        corner_embedding[:, 1, :] += self.point_embeddings[3].weight
+        return corner_embedding
+
+    def _embed_masks(self, masks: torch.Tensor) -> torch.Tensor:
+        """Embeds mask inputs."""
+        mask_embedding = self.mask_downscaling(masks)
+        return mask_embedding
+
+    def _get_batch_size(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> int:
+        """
+        Gets the batch size of the output given the batch size of the input prompts.
+        """
+        if points is not None:
+            return points[0].shape[0]
+        elif boxes is not None:
+            return boxes.shape[0]
+        elif masks is not None:
+            return masks.shape[0]
+        else:
+            return 1
+
+    def _get_device(self) -> torch.device:
+        return self.point_embeddings[0].weight.device
+
+    def forward(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Embeds different types of prompts, returning both sparse and dense
+        embeddings.
+
+        Arguments:
+          points (tuple(torch.Tensor, torch.Tensor) or none): point coordinates
+            and labels to embed.
+          boxes (torch.Tensor or none): boxes to embed
+          masks (torch.Tensor or none): masks to embed
+
+        Returns:
+          torch.Tensor: sparse embeddings for the points and boxes, with shape
+            BxNx(embed_dim), where N is determined by the number of input points
+            and boxes.
+          torch.Tensor: dense embeddings for the masks, in the shape
+            Bx(embed_dim)x(embed_H)x(embed_W)
+        """
+        bs = self._get_batch_size(points, boxes, masks)
+        sparse_embeddings = torch.empty(
+            (bs, 0, self.embed_dim), device=self._get_device()
+        )
+        if points is not None:
+            coords, labels = points
+            point_embeddings = self._embed_points(coords, labels, pad=(boxes is None))
+            sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=1)
+        if boxes is not None:
+            box_embeddings = self._embed_boxes(boxes)
+            sparse_embeddings = torch.cat([sparse_embeddings, box_embeddings], dim=1)
+
+        if masks is not None:
+            dense_embeddings = self._embed_masks(masks)
+        else:
+            dense_embeddings = self.no_mask_embed.weight.reshape(1, -1, 1, 1).expand(
+                bs, -1, self.image_embedding_size[0], self.image_embedding_size[1]
+            )
+
+        return sparse_embeddings, dense_embeddings

custom_nodes/comfyui-segment-anything-2/sam2/modeling/sam/transformer.py

@@ -0,0 +1,347 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+import warnings
+from functools import partial
+from typing import Tuple, Type
+
+import torch
+import torch.nn.functional as F
+from torch import nn, Tensor
+
+from ....sam2.modeling.position_encoding import apply_rotary_enc, compute_axial_cis
+from ....sam2.modeling.sam2_utils import MLP
+
+from ....sam2.utils.misc import get_sdpa_settings
+OLD_GPU, USE_FLASH_ATTN, MATH_KERNEL_ON = get_sdpa_settings()
+
+try:
+    from torch.nn.attention import SDPBackend, sdpa_kernel
+    backends = []
+    if USE_FLASH_ATTN:
+        backends.append(SDPBackend.FLASH_ATTENTION)
+    if MATH_KERNEL_ON:
+        backends.append(SDPBackend.MATH)
+    if OLD_GPU:
+        backends.append(SDPBackend.EFFICIENT_ATTENTION)
+    OLD_TORCH = False
+except:
+    OLD_TORCH = True
+
+warnings.simplefilter(action="ignore", category=FutureWarning)
+
+class TwoWayTransformer(nn.Module):
+    def __init__(
+        self,
+        depth: int,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+    ) -> None:
+        """
+        A transformer decoder that attends to an input image using
+        queries whose positional embedding is supplied.
+
+        Args:
+          depth (int): number of layers in the transformer
+          embedding_dim (int): the channel dimension for the input embeddings
+          num_heads (int): the number of heads for multihead attention. Must
+            divide embedding_dim
+          mlp_dim (int): the channel dimension internal to the MLP block
+          activation (nn.Module): the activation to use in the MLP block
+        """
+        super().__init__()
+        self.depth = depth
+        self.embedding_dim = embedding_dim
+        self.num_heads = num_heads
+        self.mlp_dim = mlp_dim
+        self.layers = nn.ModuleList()
+
+        for i in range(depth):
+            self.layers.append(
+                TwoWayAttentionBlock(
+                    embedding_dim=embedding_dim,
+                    num_heads=num_heads,
+                    mlp_dim=mlp_dim,
+                    activation=activation,
+                    attention_downsample_rate=attention_downsample_rate,
+                    skip_first_layer_pe=(i == 0),
+                )
+            )
+
+        self.final_attn_token_to_image = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+        self.norm_final_attn = nn.LayerNorm(embedding_dim)
+
+    def forward(
+        self,
+        image_embedding: Tensor,
+        image_pe: Tensor,
+        point_embedding: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+          image_embedding (torch.Tensor): image to attend to. Should be shape
+            B x embedding_dim x h x w for any h and w.
+          image_pe (torch.Tensor): the positional encoding to add to the image. Must
+            have the same shape as image_embedding.
+          point_embedding (torch.Tensor): the embedding to add to the query points.
+            Must have shape B x N_points x embedding_dim for any N_points.
+
+        Returns:
+          torch.Tensor: the processed point_embedding
+          torch.Tensor: the processed image_embedding
+        """
+        # BxCxHxW -> BxHWxC == B x N_image_tokens x C
+        bs, c, h, w = image_embedding.shape
+        image_embedding = image_embedding.flatten(2).permute(0, 2, 1)
+        image_pe = image_pe.flatten(2).permute(0, 2, 1)
+
+        # Prepare queries
+        queries = point_embedding
+        keys = image_embedding
+
+        # Apply transformer blocks and final layernorm
+        for layer in self.layers:
+            queries, keys = layer(
+                queries=queries,
+                keys=keys,
+                query_pe=point_embedding,
+                key_pe=image_pe,
+            )
+
+        # Apply the final attention layer from the points to the image
+        q = queries + point_embedding
+        k = keys + image_pe
+        attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm_final_attn(queries)
+
+        return queries, keys
+
+
+class TwoWayAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int = 2048,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+        skip_first_layer_pe: bool = False,
+    ) -> None:
+        """
+        A transformer block with four layers: (1) self-attention of sparse
+        inputs, (2) cross attention of sparse inputs to dense inputs, (3) mlp
+        block on sparse inputs, and (4) cross attention of dense inputs to sparse
+        inputs.
+
+        Arguments:
+          embedding_dim (int): the channel dimension of the embeddings
+          num_heads (int): the number of heads in the attention layers
+          mlp_dim (int): the hidden dimension of the mlp block
+          activation (nn.Module): the activation of the mlp block
+          skip_first_layer_pe (bool): skip the PE on the first layer
+        """
+        super().__init__()
+        self.self_attn = Attention(embedding_dim, num_heads)
+        self.norm1 = nn.LayerNorm(embedding_dim)
+
+        self.cross_attn_token_to_image = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+        self.norm2 = nn.LayerNorm(embedding_dim)
+
+        self.mlp = MLP(
+            embedding_dim, mlp_dim, embedding_dim, num_layers=2, activation=activation
+        )
+        self.norm3 = nn.LayerNorm(embedding_dim)
+
+        self.norm4 = nn.LayerNorm(embedding_dim)
+        self.cross_attn_image_to_token = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+
+        self.skip_first_layer_pe = skip_first_layer_pe
+
+    def forward(
+        self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor
+    ) -> Tuple[Tensor, Tensor]:
+        # Self attention block
+        if self.skip_first_layer_pe:
+            queries = self.self_attn(q=queries, k=queries, v=queries)
+        else:
+            q = queries + query_pe
+            attn_out = self.self_attn(q=q, k=q, v=queries)
+            queries = queries + attn_out
+        queries = self.norm1(queries)
+
+        # Cross attention block, tokens attending to image embedding
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm2(queries)
+
+        # MLP block
+        mlp_out = self.mlp(queries)
+        queries = queries + mlp_out
+        queries = self.norm3(queries)
+
+        # Cross attention block, image embedding attending to tokens
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)
+        keys = keys + attn_out
+        keys = self.norm4(keys)
+
+        return queries, keys
+
+
+class Attention(nn.Module):
+    """
+    An attention layer that allows for downscaling the size of the embedding
+    after projection to queries, keys, and values.
+    """
+
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        downsample_rate: int = 1,
+        dropout: float = 0.0,
+        kv_in_dim: int = None,
+    ) -> None:
+        super().__init__()
+        self.embedding_dim = embedding_dim
+        self.kv_in_dim = kv_in_dim if kv_in_dim is not None else embedding_dim
+        self.internal_dim = embedding_dim // downsample_rate
+        self.num_heads = num_heads
+        assert (
+            self.internal_dim % num_heads == 0
+        ), "num_heads must divide embedding_dim."
+
+        self.q_proj = nn.Linear(embedding_dim, self.internal_dim)
+        self.k_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
+        self.v_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
+        self.out_proj = nn.Linear(self.internal_dim, embedding_dim)
+
+        self.dropout_p = dropout
+
+    def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:
+        b, n, c = x.shape
+        x = x.reshape(b, n, num_heads, c // num_heads)
+        return x.transpose(1, 2)  # B x N_heads x N_tokens x C_per_head
+
+    def _recombine_heads(self, x: Tensor) -> Tensor:
+        b, n_heads, n_tokens, c_per_head = x.shape
+        x = x.transpose(1, 2)
+        return x.reshape(b, n_tokens, n_heads * c_per_head)  # B x N_tokens x C
+
+    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:
+        # Input projections
+        q = self.q_proj(q)
+        k = self.k_proj(k)
+        v = self.v_proj(v)
+
+        # Separate into heads
+        q = self._separate_heads(q, self.num_heads)
+        k = self._separate_heads(k, self.num_heads)
+        v = self._separate_heads(v, self.num_heads)
+
+        dropout_p = self.dropout_p if self.training else 0.0
+        # Attention
+        if not OLD_TORCH:
+            if not MATH_KERNEL_ON and OLD_GPU and dropout_p > 0.0:
+                backends.append(SDPBackend.MATH)
+            with sdpa_kernel(backends):
+                out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+        else:
+            with torch.backends.cuda.sdp_kernel(
+                enable_flash=USE_FLASH_ATTN,
+                enable_math=(OLD_GPU and dropout_p > 0.0) or MATH_KERNEL_ON,
+                enable_mem_efficient=OLD_GPU,
+            ):
+                out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+        out = self._recombine_heads(out)
+        out = self.out_proj(out)
+
+        return out
+
+
+class RoPEAttention(Attention):
+    """Attention with rotary position encoding."""
+
+    def __init__(
+        self,
+        *args,
+        rope_theta=10000.0,
+        # whether to repeat q rope to match k length
+        # this is needed for cross-attention to memories
+        rope_k_repeat=False,
+        feat_sizes=(32, 32),  # [w, h] for stride 16 feats at 512 resolution
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+
+        self.compute_cis = partial(
+            compute_axial_cis, dim=self.internal_dim // self.num_heads, theta=rope_theta
+        )
+        freqs_cis = self.compute_cis(end_x=feat_sizes[0], end_y=feat_sizes[1])
+        self.freqs_cis = freqs_cis
+        self.rope_k_repeat = rope_k_repeat
+
+    def forward(
+        self, q: Tensor, k: Tensor, v: Tensor, num_k_exclude_rope: int = 0
+    ) -> Tensor:
+        # Input projections
+        q = self.q_proj(q)
+        k = self.k_proj(k)
+        v = self.v_proj(v)
+
+        # Separate into heads
+        q = self._separate_heads(q, self.num_heads)
+        k = self._separate_heads(k, self.num_heads)
+        v = self._separate_heads(v, self.num_heads)
+
+        # Apply rotary position encoding
+        w = h = math.sqrt(q.shape[-2])
+        self.freqs_cis = self.freqs_cis.to(q.device)
+        if self.freqs_cis.shape[0] != q.shape[-2]:
+            self.freqs_cis = self.compute_cis(end_x=w, end_y=h).to(q.device)
+        if q.shape[-2] != k.shape[-2]:
+            assert self.rope_k_repeat
+
+        num_k_rope = k.size(-2) - num_k_exclude_rope
+        q, k[:, :, :num_k_rope] = apply_rotary_enc(
+            q,
+            k[:, :, :num_k_rope],
+            freqs_cis=self.freqs_cis,
+            repeat_freqs_k=self.rope_k_repeat,
+        )
+
+        dropout_p = self.dropout_p if self.training else 0.0
+        # Attention
+        if not OLD_TORCH:
+            if not MATH_KERNEL_ON and OLD_GPU and dropout_p > 0.0:
+                backends.append(SDPBackend.MATH)
+            with sdpa_kernel(backends):
+                out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+        else:
+            with torch.backends.cuda.sdp_kernel(
+                enable_flash=USE_FLASH_ATTN,
+                enable_math=(OLD_GPU and dropout_p > 0.0) or MATH_KERNEL_ON,
+                enable_mem_efficient=OLD_GPU,
+            ):
+                out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+        out = self._recombine_heads(out)
+        out = self.out_proj(out)
+
+        return out

custom_nodes/comfyui-segment-anything-2/sam2/modeling/sam2_base.py

@@ -0,0 +1,907 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.distributed
+import torch.nn.functional as F
+
+from torch.nn.init import trunc_normal_
+
+from ...sam2.modeling.sam.mask_decoder import MaskDecoder
+from ...sam2.modeling.sam.prompt_encoder import PromptEncoder
+from ...sam2.modeling.sam.transformer import TwoWayTransformer
+from ...sam2.modeling.sam2_utils import get_1d_sine_pe, MLP, select_closest_cond_frames
+
+# a large negative value as a placeholder score for missing objects
+NO_OBJ_SCORE = -1024.0
+
+
+class SAM2Base(torch.nn.Module):
+    def __init__(
+        self,
+        image_encoder,
+        memory_attention,
+        memory_encoder,
+        num_maskmem=7,  # default 1 input frame + 6 previous frames
+        image_size=512,
+        backbone_stride=16,  # stride of the image backbone output
+        sigmoid_scale_for_mem_enc=1.0,  # scale factor for mask sigmoid prob
+        sigmoid_bias_for_mem_enc=0.0,  # bias factor for mask sigmoid prob
+        # During evaluation, whether to binarize the sigmoid mask logits on interacted frames with clicks
+        binarize_mask_from_pts_for_mem_enc=False,
+        use_mask_input_as_output_without_sam=False,  # on frames with mask input, whether to directly output the input mask without using a SAM prompt encoder + mask decoder
+        # The maximum number of conditioning frames to participate in the memory attention (-1 means no limit; if there are more conditioning frames than this limit,
+        # we only cross-attend to the temporally closest `max_cond_frames_in_attn` conditioning frames in the encoder when tracking each frame). This gives the model
+        # a temporal locality when handling a large number of annotated frames (since closer frames should be more important) and also avoids GPU OOM.
+        max_cond_frames_in_attn=-1,
+        # on the first frame, whether to directly add the no-memory embedding to the image feature
+        # (instead of using the transformer encoder)
+        directly_add_no_mem_embed=False,
+        # whether to use high-resolution feature maps in the SAM mask decoder
+        use_high_res_features_in_sam=False,
+        # whether to output multiple (3) masks for the first click on initial conditioning frames
+        multimask_output_in_sam=False,
+        # the minimum and maximum number of clicks to use multimask_output_in_sam (only relevant when `multimask_output_in_sam=True`;
+        # default is 1 for both, meaning that only the first click gives multimask output; also note that a box counts as two points)
+        multimask_min_pt_num=1,
+        multimask_max_pt_num=1,
+        # whether to also use multimask output for tracking (not just for the first click on initial conditioning frames; only relevant when `multimask_output_in_sam=True`)
+        multimask_output_for_tracking=False,
+        # Whether to use multimask tokens for obj ptr; Only relevant when both
+        # use_obj_ptrs_in_encoder=True and multimask_output_for_tracking=True
+        use_multimask_token_for_obj_ptr: bool = False,
+        # whether to use sigmoid to restrict ious prediction to [0-1]
+        iou_prediction_use_sigmoid=False,
+        # The memory bank's temporal stride during evaluation (i.e. the `r` parameter in XMem and Cutie; XMem and Cutie use r=5).
+        # For r>1, the (self.num_maskmem - 1) non-conditioning memory frames consist of
+        # (self.num_maskmem - 2) nearest frames from every r-th frames, plus the last frame.
+        memory_temporal_stride_for_eval=1,
+        # whether to apply non-overlapping constraints on the object masks in the memory encoder during evaluation (to avoid/alleviate superposing masks)
+        non_overlap_masks_for_mem_enc=False,
+        # whether to cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+        use_obj_ptrs_in_encoder=False,
+        # the maximum number of object pointers from other frames in encoder cross attention (only relevant when `use_obj_ptrs_in_encoder=True`)
+        max_obj_ptrs_in_encoder=16,
+        # whether to add temporal positional encoding to the object pointers in the encoder (only relevant when `use_obj_ptrs_in_encoder=True`)
+        add_tpos_enc_to_obj_ptrs=True,
+        # whether to add an extra linear projection layer for the temporal positional encoding in the object pointers to avoid potential interference
+        # with spatial positional encoding (only relevant when both `use_obj_ptrs_in_encoder=True` and `add_tpos_enc_to_obj_ptrs=True`)
+        proj_tpos_enc_in_obj_ptrs=False,
+        # whether to use signed distance (instead of unsigned absolute distance) in the temporal positional encoding in the object pointers
+        # (only relevant when both `use_obj_ptrs_in_encoder=True` and `add_tpos_enc_to_obj_ptrs=True`)
+        use_signed_tpos_enc_to_obj_ptrs=False,
+        # whether to only attend to object pointers in the past (before the current frame) in the encoder during evaluation
+        # (only relevant when `use_obj_ptrs_in_encoder=True`; this might avoid pointer information too far in the future to distract the initial tracking)
+        only_obj_ptrs_in_the_past_for_eval=False,
+        # Whether to predict if there is an object in the frame
+        pred_obj_scores: bool = False,
+        # Whether to use an MLP to predict object scores
+        pred_obj_scores_mlp: bool = False,
+        # Only relevant if pred_obj_scores=True and use_obj_ptrs_in_encoder=True;
+        # Whether to have a fixed no obj pointer when there is no object present
+        # or to use it as an additive embedding with obj_ptr produced by decoder
+        fixed_no_obj_ptr: bool = False,
+        # Soft no object, i.e. mix in no_obj_ptr softly,
+        # hope to make recovery easier if there is a mistake and mitigate accumulation of errors
+        soft_no_obj_ptr: bool = False,
+        use_mlp_for_obj_ptr_proj: bool = False,
+        # add no obj embedding to spatial frames
+        no_obj_embed_spatial: bool = False,
+        # extra arguments used to construct the SAM mask decoder; if not None, it should be a dict of kwargs to be passed into `MaskDecoder` class.
+        sam_mask_decoder_extra_args=None,
+        compile_image_encoder: bool = False,
+    ):
+        super().__init__()
+
+        # Part 1: the image backbone
+        self.image_encoder = image_encoder
+        # Use level 0, 1, 2 for high-res setting, or just level 2 for the default setting
+        self.use_high_res_features_in_sam = use_high_res_features_in_sam
+        self.num_feature_levels = 3 if use_high_res_features_in_sam else 1
+        self.use_obj_ptrs_in_encoder = use_obj_ptrs_in_encoder
+        self.max_obj_ptrs_in_encoder = max_obj_ptrs_in_encoder
+        if use_obj_ptrs_in_encoder:
+            # A conv layer to downsample the mask prompt to stride 4 (the same stride as
+            # low-res SAM mask logits) and to change its scales from 0~1 to SAM logit scale,
+            # so that it can be fed into the SAM mask decoder to generate a pointer.
+            self.mask_downsample = torch.nn.Conv2d(1, 1, kernel_size=4, stride=4)
+        self.add_tpos_enc_to_obj_ptrs = add_tpos_enc_to_obj_ptrs
+        if proj_tpos_enc_in_obj_ptrs:
+            assert add_tpos_enc_to_obj_ptrs  # these options need to be used together
+        self.proj_tpos_enc_in_obj_ptrs = proj_tpos_enc_in_obj_ptrs
+        self.use_signed_tpos_enc_to_obj_ptrs = use_signed_tpos_enc_to_obj_ptrs
+        self.only_obj_ptrs_in_the_past_for_eval = only_obj_ptrs_in_the_past_for_eval
+
+        # Part 2: memory attention to condition current frame's visual features
+        # with memories (and obj ptrs) from past frames
+        self.memory_attention = memory_attention
+        self.hidden_dim = image_encoder.neck.d_model
+
+        # Part 3: memory encoder for the previous frame's outputs
+        self.memory_encoder = memory_encoder
+        self.mem_dim = self.hidden_dim
+        if hasattr(self.memory_encoder, "out_proj") and hasattr(
+            self.memory_encoder.out_proj, "weight"
+        ):
+            # if there is compression of memories along channel dim
+            self.mem_dim = self.memory_encoder.out_proj.weight.shape[0]
+        self.num_maskmem = num_maskmem  # Number of memories accessible
+        # Temporal encoding of the memories
+        self.maskmem_tpos_enc = torch.nn.Parameter(
+            torch.zeros(num_maskmem, 1, 1, self.mem_dim)
+        )
+        trunc_normal_(self.maskmem_tpos_enc, std=0.02)
+        # a single token to indicate no memory embedding from previous frames
+        self.no_mem_embed = torch.nn.Parameter(torch.zeros(1, 1, self.hidden_dim))
+        self.no_mem_pos_enc = torch.nn.Parameter(torch.zeros(1, 1, self.hidden_dim))
+        trunc_normal_(self.no_mem_embed, std=0.02)
+        trunc_normal_(self.no_mem_pos_enc, std=0.02)
+        self.directly_add_no_mem_embed = directly_add_no_mem_embed
+        # Apply sigmoid to the output raw mask logits (to turn them from
+        # range (-inf, +inf) to range (0, 1)) before feeding them into the memory encoder
+        self.sigmoid_scale_for_mem_enc = sigmoid_scale_for_mem_enc
+        self.sigmoid_bias_for_mem_enc = sigmoid_bias_for_mem_enc
+        self.binarize_mask_from_pts_for_mem_enc = binarize_mask_from_pts_for_mem_enc
+        self.non_overlap_masks_for_mem_enc = non_overlap_masks_for_mem_enc
+        self.memory_temporal_stride_for_eval = memory_temporal_stride_for_eval
+        # On frames with mask input, whether to directly output the input mask without
+        # using a SAM prompt encoder + mask decoder
+        self.use_mask_input_as_output_without_sam = use_mask_input_as_output_without_sam
+        self.multimask_output_in_sam = multimask_output_in_sam
+        self.multimask_min_pt_num = multimask_min_pt_num
+        self.multimask_max_pt_num = multimask_max_pt_num
+        self.multimask_output_for_tracking = multimask_output_for_tracking
+        self.use_multimask_token_for_obj_ptr = use_multimask_token_for_obj_ptr
+        self.iou_prediction_use_sigmoid = iou_prediction_use_sigmoid
+
+        # Part 4: SAM-style prompt encoder (for both mask and point inputs)
+        # and SAM-style mask decoder for the final mask output
+        self.image_size = image_size
+        self.backbone_stride = backbone_stride
+        self.sam_mask_decoder_extra_args = sam_mask_decoder_extra_args
+        self.pred_obj_scores = pred_obj_scores
+        self.pred_obj_scores_mlp = pred_obj_scores_mlp
+        self.fixed_no_obj_ptr = fixed_no_obj_ptr
+        self.soft_no_obj_ptr = soft_no_obj_ptr
+        if self.fixed_no_obj_ptr:
+            assert self.pred_obj_scores
+            assert self.use_obj_ptrs_in_encoder
+        if self.pred_obj_scores and self.use_obj_ptrs_in_encoder:
+            self.no_obj_ptr = torch.nn.Parameter(torch.zeros(1, self.hidden_dim))
+            trunc_normal_(self.no_obj_ptr, std=0.02)
+        self.use_mlp_for_obj_ptr_proj = use_mlp_for_obj_ptr_proj
+        self.no_obj_embed_spatial = None
+        if no_obj_embed_spatial:
+            self.no_obj_embed_spatial = torch.nn.Parameter(torch.zeros(1, self.mem_dim))
+            trunc_normal_(self.no_obj_embed_spatial, std=0.02)
+
+        self._build_sam_heads()
+        self.max_cond_frames_in_attn = max_cond_frames_in_attn
+
+        # Model compilation
+        if compile_image_encoder:
+            # Compile the forward function (not the full module) to allow loading checkpoints.
+            print(
+                "Image encoder compilation is enabled. First forward pass will be slow."
+            )
+            self.image_encoder.forward = torch.compile(
+                self.image_encoder.forward,
+                mode="max-autotune",
+                fullgraph=True,
+                dynamic=False,
+            )
+
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    def forward(self, *args, **kwargs):
+        raise NotImplementedError(
+            "Please use the corresponding methods in SAM2VideoPredictor for inference or SAM2Train for training/fine-tuning"
+            "See notebooks/video_predictor_example.ipynb for an inference example."
+        )
+
+    def _build_sam_heads(self):
+        """Build SAM-style prompt encoder and mask decoder."""
+        self.sam_prompt_embed_dim = self.hidden_dim
+        self.sam_image_embedding_size = self.image_size // self.backbone_stride
+
+        # build PromptEncoder and MaskDecoder from SAM
+        # (their hyperparameters like `mask_in_chans=16` are from SAM code)
+        self.sam_prompt_encoder = PromptEncoder(
+            embed_dim=self.sam_prompt_embed_dim,
+            image_embedding_size=(
+                self.sam_image_embedding_size,
+                self.sam_image_embedding_size,
+            ),
+            input_image_size=(self.image_size, self.image_size),
+            mask_in_chans=16,
+        )
+        self.sam_mask_decoder = MaskDecoder(
+            num_multimask_outputs=3,
+            transformer=TwoWayTransformer(
+                depth=2,
+                embedding_dim=self.sam_prompt_embed_dim,
+                mlp_dim=2048,
+                num_heads=8,
+            ),
+            transformer_dim=self.sam_prompt_embed_dim,
+            iou_head_depth=3,
+            iou_head_hidden_dim=256,
+            use_high_res_features=self.use_high_res_features_in_sam,
+            iou_prediction_use_sigmoid=self.iou_prediction_use_sigmoid,
+            pred_obj_scores=self.pred_obj_scores,
+            pred_obj_scores_mlp=self.pred_obj_scores_mlp,
+            use_multimask_token_for_obj_ptr=self.use_multimask_token_for_obj_ptr,
+            **(self.sam_mask_decoder_extra_args or {}),
+        )
+        if self.use_obj_ptrs_in_encoder:
+            # a linear projection on SAM output tokens to turn them into object pointers
+            self.obj_ptr_proj = torch.nn.Linear(self.hidden_dim, self.hidden_dim)
+            if self.use_mlp_for_obj_ptr_proj:
+                self.obj_ptr_proj = MLP(
+                    self.hidden_dim, self.hidden_dim, self.hidden_dim, 3
+                )
+        else:
+            self.obj_ptr_proj = torch.nn.Identity()
+        if self.proj_tpos_enc_in_obj_ptrs:
+            # a linear projection on temporal positional encoding in object pointers to
+            # avoid potential interference with spatial positional encoding
+            self.obj_ptr_tpos_proj = torch.nn.Linear(self.hidden_dim, self.mem_dim)
+        else:
+            self.obj_ptr_tpos_proj = torch.nn.Identity()
+
+    def _forward_sam_heads(
+        self,
+        backbone_features,
+        point_inputs=None,
+        mask_inputs=None,
+        high_res_features=None,
+        multimask_output=False,
+    ):
+        """
+        Forward SAM prompt encoders and mask heads.
+
+        Inputs:
+        - backbone_features: image features of [B, C, H, W] shape
+        - point_inputs: a dictionary with "point_coords" and "point_labels", where
+          1) "point_coords" has [B, P, 2] shape and float32 dtype and contains the
+             absolute pixel-unit coordinate in (x, y) format of the P input points
+          2) "point_labels" has shape [B, P] and int32 dtype, where 1 means
+             positive clicks, 0 means negative clicks, and -1 means padding
+        - mask_inputs: a mask of [B, 1, H*16, W*16] shape, float or bool, with the
+          same spatial size as the image.
+        - high_res_features: either 1) None or 2) or a list of length 2 containing
+          two feature maps of [B, C, 4*H, 4*W] and [B, C, 2*H, 2*W] shapes respectively,
+          which will be used as high-resolution feature maps for SAM decoder.
+        - multimask_output: if it's True, we output 3 candidate masks and their 3
+          corresponding IoU estimates, and if it's False, we output only 1 mask and
+          its corresponding IoU estimate.
+
+        Outputs:
+        - low_res_multimasks: [B, M, H*4, W*4] shape (where M = 3 if
+          `multimask_output=True` and M = 1 if `multimask_output=False`), the SAM
+          output mask logits (before sigmoid) for the low-resolution masks, with 4x
+          the resolution (1/4 stride) of the input backbone_features.
+        - high_res_multimasks: [B, M, H*16, W*16] shape (where M = 3
+          if `multimask_output=True` and M = 1 if `multimask_output=False`),
+          upsampled from the low-resolution masks, with shape size as the image
+          (stride is 1 pixel).
+        - ious, [B, M] shape, where (where M = 3 if `multimask_output=True` and M = 1
+          if `multimask_output=False`), the estimated IoU of each output mask.
+        - low_res_masks: [B, 1, H*4, W*4] shape, the best mask in `low_res_multimasks`.
+          If `multimask_output=True`, it's the mask with the highest IoU estimate.
+          If `multimask_output=False`, it's the same as `low_res_multimasks`.
+        - high_res_masks: [B, 1, H*16, W*16] shape, the best mask in `high_res_multimasks`.
+          If `multimask_output=True`, it's the mask with the highest IoU estimate.
+          If `multimask_output=False`, it's the same as `high_res_multimasks`.
+        - obj_ptr: [B, C] shape, the object pointer vector for the output mask, extracted
+          based on the output token from the SAM mask decoder.
+        """
+        B = backbone_features.size(0)
+        device = backbone_features.device
+        assert backbone_features.size(1) == self.sam_prompt_embed_dim
+        assert backbone_features.size(2) == self.sam_image_embedding_size
+        assert backbone_features.size(3) == self.sam_image_embedding_size
+
+        # a) Handle point prompts
+        if point_inputs is not None:
+            sam_point_coords = point_inputs["point_coords"]
+            sam_point_labels = point_inputs["point_labels"]
+            assert sam_point_coords.size(0) == B and sam_point_labels.size(0) == B
+        else:
+            # If no points are provide, pad with an empty point (with label -1)
+            sam_point_coords = torch.zeros(B, 1, 2, device=device)
+            sam_point_labels = -torch.ones(B, 1, dtype=torch.int32, device=device)
+
+        # b) Handle mask prompts
+        if mask_inputs is not None:
+            # If mask_inputs is provided, downsize it into low-res mask input if needed
+            # and feed it as a dense mask prompt into the SAM mask encoder
+            assert len(mask_inputs.shape) == 4 and mask_inputs.shape[:2] == (B, 1)
+            if mask_inputs.shape[-2:] != self.sam_prompt_encoder.mask_input_size:
+                sam_mask_prompt = F.interpolate(
+                    mask_inputs.float(),
+                    size=self.sam_prompt_encoder.mask_input_size,
+                    align_corners=False,
+                    mode="bilinear",
+                    antialias=True,  # use antialias for downsampling
+                )
+            else:
+                sam_mask_prompt = mask_inputs
+        else:
+            # Otherwise, simply feed None (and SAM's prompt encoder will add
+            # a learned `no_mask_embed` to indicate no mask input in this case).
+            sam_mask_prompt = None
+
+        sparse_embeddings, dense_embeddings = self.sam_prompt_encoder(
+            points=(sam_point_coords, sam_point_labels),
+            boxes=None,
+            masks=sam_mask_prompt,
+        )
+        (
+            low_res_multimasks,
+            ious,
+            sam_output_tokens,
+            object_score_logits,
+        ) = self.sam_mask_decoder(
+            image_embeddings=backbone_features,
+            image_pe=self.sam_prompt_encoder.get_dense_pe(),
+            sparse_prompt_embeddings=sparse_embeddings,
+            dense_prompt_embeddings=dense_embeddings,
+            multimask_output=multimask_output,
+            repeat_image=False,  # the image is already batched
+            high_res_features=high_res_features,
+        )
+        if self.pred_obj_scores:
+            is_obj_appearing = object_score_logits > 0
+
+            # Mask used for spatial memories is always a *hard* choice between obj and no obj,
+            # consistent with the actual mask prediction
+            low_res_multimasks = torch.where(
+                is_obj_appearing[:, None, None],
+                low_res_multimasks,
+                NO_OBJ_SCORE,
+            )
+
+        # convert masks from possibly bfloat16 (or float16) to float32
+        # (older PyTorch versions before 2.1 don't support `interpolate` on bf16)
+        low_res_multimasks = low_res_multimasks.float()
+        high_res_multimasks = F.interpolate(
+            low_res_multimasks,
+            size=(self.image_size, self.image_size),
+            mode="bilinear",
+            align_corners=False,
+        )
+
+        sam_output_token = sam_output_tokens[:, 0]
+        if multimask_output:
+            # take the best mask prediction (with the highest IoU estimation)
+            best_iou_inds = torch.argmax(ious, dim=-1)
+            batch_inds = torch.arange(B, device=device)
+            low_res_masks = low_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
+            high_res_masks = high_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
+            if sam_output_tokens.size(1) > 1:
+                sam_output_token = sam_output_tokens[batch_inds, best_iou_inds]
+        else:
+            low_res_masks, high_res_masks = low_res_multimasks, high_res_multimasks
+
+        # Extract object pointer from the SAM output token (with occlusion handling)
+        obj_ptr = self.obj_ptr_proj(sam_output_token)
+        if self.pred_obj_scores:
+            # Allow *soft* no obj ptr, unlike for masks
+            if self.soft_no_obj_ptr:
+                lambda_is_obj_appearing = object_score_logits.sigmoid()
+            else:
+                lambda_is_obj_appearing = is_obj_appearing.float()
+
+            if self.fixed_no_obj_ptr:
+                obj_ptr = lambda_is_obj_appearing * obj_ptr
+            obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr
+
+        return (
+            low_res_multimasks,
+            high_res_multimasks,
+            ious,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            object_score_logits,
+        )
+
+    def _use_mask_as_output(self, backbone_features, high_res_features, mask_inputs):
+        """
+        Directly turn binary `mask_inputs` into a output mask logits without using SAM.
+        (same input and output shapes as in _forward_sam_heads above).
+        """
+        # Use -10/+10 as logits for neg/pos pixels (very close to 0/1 in prob after sigmoid).
+        out_scale, out_bias = 20.0, -10.0  # sigmoid(-10.0)=4.5398e-05
+        mask_inputs_float = mask_inputs.float()
+        high_res_masks = mask_inputs_float * out_scale + out_bias
+        low_res_masks = F.interpolate(
+            high_res_masks,
+            size=(high_res_masks.size(-2) // 4, high_res_masks.size(-1) // 4),
+            align_corners=False,
+            mode="bilinear",
+            antialias=True,  # use antialias for downsampling
+        )
+        # a dummy IoU prediction of all 1's under mask input
+        ious = mask_inputs.new_ones(mask_inputs.size(0), 1).float()
+        if not self.use_obj_ptrs_in_encoder:
+            # all zeros as a dummy object pointer (of shape [B, C])
+            obj_ptr = torch.zeros(
+                mask_inputs.size(0), self.hidden_dim, device=mask_inputs.device
+            )
+        else:
+            # produce an object pointer using the SAM decoder from the mask input
+            _, _, _, _, _, obj_ptr, _ = self._forward_sam_heads(
+                backbone_features=backbone_features,
+                mask_inputs=self.mask_downsample(mask_inputs_float),
+                high_res_features=high_res_features,
+            )
+        # In this method, we are treating mask_input as output, e.g. using it directly to create spatial mem;
+        # Below, we follow the same design axiom to use mask_input to decide if obj appears or not instead of relying
+        # on the object_scores from the SAM decoder.
+        is_obj_appearing = torch.any(mask_inputs.flatten(1).float() > 0.0, dim=1)
+        is_obj_appearing = is_obj_appearing[..., None]
+        lambda_is_obj_appearing = is_obj_appearing.float()
+        object_score_logits = out_scale * lambda_is_obj_appearing + out_bias
+        if self.pred_obj_scores:
+            if self.fixed_no_obj_ptr:
+                obj_ptr = lambda_is_obj_appearing * obj_ptr
+            obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr
+
+        return (
+            low_res_masks,
+            high_res_masks,
+            ious,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            object_score_logits,
+        )
+
+    def forward_image(self, img_batch: torch.Tensor):
+        """Get the image feature on the input batch."""
+        backbone_out = self.image_encoder(img_batch)
+        if self.use_high_res_features_in_sam:
+            # precompute projected level 0 and level 1 features in SAM decoder
+            # to avoid running it again on every SAM click
+            backbone_out["backbone_fpn"][0] = self.sam_mask_decoder.conv_s0(
+                backbone_out["backbone_fpn"][0]
+            )
+            backbone_out["backbone_fpn"][1] = self.sam_mask_decoder.conv_s1(
+                backbone_out["backbone_fpn"][1]
+            )
+        return backbone_out
+
+    def _prepare_backbone_features(self, backbone_out):
+        """Prepare and flatten visual features."""
+        backbone_out = backbone_out.copy()
+        assert len(backbone_out["backbone_fpn"]) == len(backbone_out["vision_pos_enc"])
+        assert len(backbone_out["backbone_fpn"]) >= self.num_feature_levels
+
+        feature_maps = backbone_out["backbone_fpn"][-self.num_feature_levels :]
+        vision_pos_embeds = backbone_out["vision_pos_enc"][-self.num_feature_levels :]
+
+        feat_sizes = [(x.shape[-2], x.shape[-1]) for x in vision_pos_embeds]
+        # flatten NxCxHxW to HWxNxC
+        vision_feats = [x.flatten(2).permute(2, 0, 1) for x in feature_maps]
+        vision_pos_embeds = [x.flatten(2).permute(2, 0, 1) for x in vision_pos_embeds]
+
+        return backbone_out, vision_feats, vision_pos_embeds, feat_sizes
+
+    def _prepare_memory_conditioned_features(
+        self,
+        frame_idx,
+        is_init_cond_frame,
+        current_vision_feats,
+        current_vision_pos_embeds,
+        feat_sizes,
+        output_dict,
+        num_frames,
+        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
+    ):
+        """Fuse the current frame's visual feature map with previous memory."""
+        B = current_vision_feats[-1].size(1)  # batch size on this frame
+        C = self.hidden_dim
+        H, W = feat_sizes[-1]  # top-level (lowest-resolution) feature size
+        device = current_vision_feats[-1].device
+        # The case of `self.num_maskmem == 0` below is primarily used for reproducing SAM on images.
+        # In this case, we skip the fusion with any memory.
+        if self.num_maskmem == 0:  # Disable memory and skip fusion
+            pix_feat = current_vision_feats[-1].permute(1, 2, 0).view(B, C, H, W)
+            return pix_feat
+
+        num_obj_ptr_tokens = 0
+        tpos_sign_mul = -1 if track_in_reverse else 1
+        # Step 1: condition the visual features of the current frame on previous memories
+        if not is_init_cond_frame:
+            # Retrieve the memories encoded with the maskmem backbone
+            to_cat_memory, to_cat_memory_pos_embed = [], []
+            # Add conditioning frames's output first (all cond frames have t_pos=0 for
+            # when getting temporal positional embedding below)
+            assert len(output_dict["cond_frame_outputs"]) > 0
+            # Select a maximum number of temporally closest cond frames for cross attention
+            cond_outputs = output_dict["cond_frame_outputs"]
+            selected_cond_outputs, unselected_cond_outputs = select_closest_cond_frames(
+                frame_idx, cond_outputs, self.max_cond_frames_in_attn
+            )
+            t_pos_and_prevs = [(0, out) for out in selected_cond_outputs.values()]
+            # Add last (self.num_maskmem - 1) frames before current frame for non-conditioning memory
+            # the earliest one has t_pos=1 and the latest one has t_pos=self.num_maskmem-1
+            # We also allow taking the memory frame non-consecutively (with stride>1), in which case
+            # we take (self.num_maskmem - 2) frames among every stride-th frames plus the last frame.
+            stride = 1 if self.training else self.memory_temporal_stride_for_eval
+            for t_pos in range(1, self.num_maskmem):
+                t_rel = self.num_maskmem - t_pos  # how many frames before current frame
+                if t_rel == 1:
+                    # for t_rel == 1, we take the last frame (regardless of r)
+                    if not track_in_reverse:
+                        # the frame immediately before this frame (i.e. frame_idx - 1)
+                        prev_frame_idx = frame_idx - t_rel
+                    else:
+                        # the frame immediately after this frame (i.e. frame_idx + 1)
+                        prev_frame_idx = frame_idx + t_rel
+                else:
+                    # for t_rel >= 2, we take the memory frame from every r-th frames
+                    if not track_in_reverse:
+                        # first find the nearest frame among every r-th frames before this frame
+                        # for r=1, this would be (frame_idx - 2)
+                        prev_frame_idx = ((frame_idx - 2) // stride) * stride
+                        # then seek further among every r-th frames
+                        prev_frame_idx = prev_frame_idx - (t_rel - 2) * stride
+                    else:
+                        # first find the nearest frame among every r-th frames after this frame
+                        # for r=1, this would be (frame_idx + 2)
+                        prev_frame_idx = -(-(frame_idx + 2) // stride) * stride
+                        # then seek further among every r-th frames
+                        prev_frame_idx = prev_frame_idx + (t_rel - 2) * stride
+                out = output_dict["non_cond_frame_outputs"].get(prev_frame_idx, None)
+                if out is None:
+                    # If an unselected conditioning frame is among the last (self.num_maskmem - 1)
+                    # frames, we still attend to it as if it's a non-conditioning frame.
+                    out = unselected_cond_outputs.get(prev_frame_idx, None)
+                t_pos_and_prevs.append((t_pos, out))
+
+            for t_pos, prev in t_pos_and_prevs:
+                if prev is None:
+                    continue  # skip padding frames
+                # "maskmem_features" might have been offloaded to CPU in demo use cases,
+                # so we load it back to GPU (it's a no-op if it's already on GPU).
+                feats = prev["maskmem_features"].to(device, non_blocking=True)
+                to_cat_memory.append(feats.flatten(2).permute(2, 0, 1))
+                # Spatial positional encoding (it might have been offloaded to CPU in eval)
+                maskmem_enc = prev["maskmem_pos_enc"][-1].to(device)
+                maskmem_enc = maskmem_enc.flatten(2).permute(2, 0, 1)
+                # Temporal positional encoding
+                maskmem_enc = (
+                    maskmem_enc + self.maskmem_tpos_enc[self.num_maskmem - t_pos - 1]
+                )
+                to_cat_memory_pos_embed.append(maskmem_enc)
+
+            # Construct the list of past object pointers
+            if self.use_obj_ptrs_in_encoder:
+                max_obj_ptrs_in_encoder = min(num_frames, self.max_obj_ptrs_in_encoder)
+                # First add those object pointers from selected conditioning frames
+                # (optionally, only include object pointers in the past during evaluation)
+                if not self.training and self.only_obj_ptrs_in_the_past_for_eval:
+                    ptr_cond_outputs = {
+                        t: out
+                        for t, out in selected_cond_outputs.items()
+                        if (t >= frame_idx if track_in_reverse else t <= frame_idx)
+                    }
+                else:
+                    ptr_cond_outputs = selected_cond_outputs
+                pos_and_ptrs = [
+                    # Temporal pos encoding contains how far away each pointer is from current frame
+                    (
+                        (
+                            (frame_idx - t) * tpos_sign_mul
+                            if self.use_signed_tpos_enc_to_obj_ptrs
+                            else abs(frame_idx - t)
+                        ),
+                        out["obj_ptr"],
+                    )
+                    for t, out in ptr_cond_outputs.items()
+                ]
+                # Add up to (max_obj_ptrs_in_encoder - 1) non-conditioning frames before current frame
+                for t_diff in range(1, max_obj_ptrs_in_encoder):
+                    t = frame_idx + t_diff if track_in_reverse else frame_idx - t_diff
+                    if t < 0 or (num_frames is not None and t >= num_frames):
+                        break
+                    out = output_dict["non_cond_frame_outputs"].get(
+                        t, unselected_cond_outputs.get(t, None)
+                    )
+                    if out is not None:
+                        pos_and_ptrs.append((t_diff, out["obj_ptr"]))
+                # If we have at least one object pointer, add them to the across attention
+                if len(pos_and_ptrs) > 0:
+                    pos_list, ptrs_list = zip(*pos_and_ptrs)
+                    # stack object pointers along dim=0 into [ptr_seq_len, B, C] shape
+                    obj_ptrs = torch.stack(ptrs_list, dim=0)
+                    # a temporal positional embedding based on how far each object pointer is from
+                    # the current frame (sine embedding normalized by the max pointer num).
+                    if self.add_tpos_enc_to_obj_ptrs:
+                        t_diff_max = max_obj_ptrs_in_encoder - 1
+                        tpos_dim = C if self.proj_tpos_enc_in_obj_ptrs else self.mem_dim
+                        obj_pos = torch.tensor(pos_list, device=device)
+                        obj_pos = get_1d_sine_pe(obj_pos / t_diff_max, dim=tpos_dim)
+                        obj_pos = self.obj_ptr_tpos_proj(obj_pos)
+                        obj_pos = obj_pos.unsqueeze(1).expand(-1, B, self.mem_dim)
+                    else:
+                        obj_pos = obj_ptrs.new_zeros(len(pos_list), B, self.mem_dim)
+                    if self.mem_dim < C:
+                        # split a pointer into (C // self.mem_dim) tokens for self.mem_dim < C
+                        obj_ptrs = obj_ptrs.reshape(
+                            -1, B, C // self.mem_dim, self.mem_dim
+                        )
+                        obj_ptrs = obj_ptrs.permute(0, 2, 1, 3).flatten(0, 1)
+                        obj_pos = obj_pos.repeat_interleave(C // self.mem_dim, dim=0)
+                    to_cat_memory.append(obj_ptrs)
+                    to_cat_memory_pos_embed.append(obj_pos)
+                    num_obj_ptr_tokens = obj_ptrs.shape[0]
+                else:
+                    num_obj_ptr_tokens = 0
+        else:
+            # for initial conditioning frames, encode them without using any previous memory
+            if self.directly_add_no_mem_embed:
+                # directly add no-mem embedding (instead of using the transformer encoder)
+                pix_feat_with_mem = current_vision_feats[-1] + self.no_mem_embed
+                pix_feat_with_mem = pix_feat_with_mem.permute(1, 2, 0).view(B, C, H, W)
+                return pix_feat_with_mem
+
+            # Use a dummy token on the first frame (to avoid empty memory input to tranformer encoder)
+            to_cat_memory = [self.no_mem_embed.expand(1, B, self.mem_dim)]
+            to_cat_memory_pos_embed = [self.no_mem_pos_enc.expand(1, B, self.mem_dim)]
+
+        # Step 2: Concatenate the memories and forward through the transformer encoder
+        memory = torch.cat(to_cat_memory, dim=0)
+        memory_pos_embed = torch.cat(to_cat_memory_pos_embed, dim=0)
+
+        pix_feat_with_mem = self.memory_attention(
+            curr=current_vision_feats,
+            curr_pos=current_vision_pos_embeds,
+            memory=memory,
+            memory_pos=memory_pos_embed,
+            num_obj_ptr_tokens=num_obj_ptr_tokens,
+        )
+        # reshape the output (HW)BC => BCHW
+        pix_feat_with_mem = pix_feat_with_mem.permute(1, 2, 0).view(B, C, H, W)
+        return pix_feat_with_mem
+
+    def _encode_new_memory(
+        self,
+        current_vision_feats,
+        feat_sizes,
+        pred_masks_high_res,
+        object_score_logits,
+        is_mask_from_pts,
+    ):
+        """Encode the current image and its prediction into a memory feature."""
+        B = current_vision_feats[-1].size(1)  # batch size on this frame
+        C = self.hidden_dim
+        H, W = feat_sizes[-1]  # top-level (lowest-resolution) feature size
+        # top-level feature, (HW)BC => BCHW
+        pix_feat = current_vision_feats[-1].permute(1, 2, 0).view(B, C, H, W)
+        if self.non_overlap_masks_for_mem_enc and not self.training:
+            # optionally, apply non-overlapping constraints to the masks (it's applied
+            # in the batch dimension and should only be used during eval, where all
+            # the objects come from the same video under batch size 1).
+            pred_masks_high_res = self._apply_non_overlapping_constraints(
+                pred_masks_high_res
+            )
+        # scale the raw mask logits with a temperature before applying sigmoid
+        binarize = self.binarize_mask_from_pts_for_mem_enc and is_mask_from_pts
+        if binarize and not self.training:
+            mask_for_mem = (pred_masks_high_res > 0).float()
+        else:
+            # apply sigmoid on the raw mask logits to turn them into range (0, 1)
+            mask_for_mem = torch.sigmoid(pred_masks_high_res)
+        # apply scale and bias terms to the sigmoid probabilities
+        if self.sigmoid_scale_for_mem_enc != 1.0:
+            mask_for_mem = mask_for_mem * self.sigmoid_scale_for_mem_enc
+        if self.sigmoid_bias_for_mem_enc != 0.0:
+            mask_for_mem = mask_for_mem + self.sigmoid_bias_for_mem_enc
+        maskmem_out = self.memory_encoder(
+            pix_feat, mask_for_mem, skip_mask_sigmoid=True  # sigmoid already applied
+        )
+        maskmem_features = maskmem_out["vision_features"]
+        maskmem_pos_enc = maskmem_out["vision_pos_enc"]
+        # add a no-object embedding to the spatial memory to indicate that the frame
+        # is predicted to be occluded (i.e. no object is appearing in the frame)
+        if self.no_obj_embed_spatial is not None:
+            is_obj_appearing = (object_score_logits > 0).float()
+            maskmem_features += (
+                1 - is_obj_appearing[..., None, None]
+            ) * self.no_obj_embed_spatial[..., None, None].expand(
+                *maskmem_features.shape
+            )
+
+        return maskmem_features, maskmem_pos_enc
+
+    def _track_step(
+        self,
+        frame_idx,
+        is_init_cond_frame,
+        current_vision_feats,
+        current_vision_pos_embeds,
+        feat_sizes,
+        point_inputs,
+        mask_inputs,
+        output_dict,
+        num_frames,
+        track_in_reverse,
+        prev_sam_mask_logits,
+    ):
+        current_out = {"point_inputs": point_inputs, "mask_inputs": mask_inputs}
+        # High-resolution feature maps for the SAM head, reshape (HW)BC => BCHW
+        if len(current_vision_feats) > 1:
+            high_res_features = [
+                x.permute(1, 2, 0).view(x.size(1), x.size(2), *s)
+                for x, s in zip(current_vision_feats[:-1], feat_sizes[:-1])
+            ]
+        else:
+            high_res_features = None
+        if mask_inputs is not None and self.use_mask_input_as_output_without_sam:
+            # When use_mask_input_as_output_without_sam=True, we directly output the mask input
+            # (see it as a GT mask) without using a SAM prompt encoder + mask decoder.
+            pix_feat = current_vision_feats[-1].permute(1, 2, 0)
+            pix_feat = pix_feat.view(-1, self.hidden_dim, *feat_sizes[-1])
+            sam_outputs = self._use_mask_as_output(
+                pix_feat, high_res_features, mask_inputs
+            )
+        else:
+            # fused the visual feature with previous memory features in the memory bank
+            pix_feat = self._prepare_memory_conditioned_features(
+                frame_idx=frame_idx,
+                is_init_cond_frame=is_init_cond_frame,
+                current_vision_feats=current_vision_feats[-1:],
+                current_vision_pos_embeds=current_vision_pos_embeds[-1:],
+                feat_sizes=feat_sizes[-1:],
+                output_dict=output_dict,
+                num_frames=num_frames,
+                track_in_reverse=track_in_reverse,
+            )
+            # apply SAM-style segmentation head
+            # here we might feed previously predicted low-res SAM mask logits into the SAM mask decoder,
+            # e.g. in demo where such logits come from earlier interaction instead of correction sampling
+            # (in this case, any `mask_inputs` shouldn't reach here as they are sent to _use_mask_as_output instead)
+            if prev_sam_mask_logits is not None:
+                assert point_inputs is not None and mask_inputs is None
+                mask_inputs = prev_sam_mask_logits
+            multimask_output = self._use_multimask(is_init_cond_frame, point_inputs)
+            sam_outputs = self._forward_sam_heads(
+                backbone_features=pix_feat,
+                point_inputs=point_inputs,
+                mask_inputs=mask_inputs,
+                high_res_features=high_res_features,
+                multimask_output=multimask_output,
+            )
+
+        return current_out, sam_outputs, high_res_features, pix_feat
+
+    def _encode_memory_in_output(
+        self,
+        current_vision_feats,
+        feat_sizes,
+        point_inputs,
+        run_mem_encoder,
+        high_res_masks,
+        object_score_logits,
+        current_out,
+    ):
+        if run_mem_encoder and self.num_maskmem > 0:
+            high_res_masks_for_mem_enc = high_res_masks
+            maskmem_features, maskmem_pos_enc = self._encode_new_memory(
+                current_vision_feats=current_vision_feats,
+                feat_sizes=feat_sizes,
+                pred_masks_high_res=high_res_masks_for_mem_enc,
+                object_score_logits=object_score_logits,
+                is_mask_from_pts=(point_inputs is not None),
+            )
+            current_out["maskmem_features"] = maskmem_features
+            current_out["maskmem_pos_enc"] = maskmem_pos_enc
+        else:
+            current_out["maskmem_features"] = None
+            current_out["maskmem_pos_enc"] = None
+
+    def track_step(
+        self,
+        frame_idx,
+        is_init_cond_frame,
+        current_vision_feats,
+        current_vision_pos_embeds,
+        feat_sizes,
+        point_inputs,
+        mask_inputs,
+        output_dict,
+        num_frames,
+        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
+        # Whether to run the memory encoder on the predicted masks. Sometimes we might want
+        # to skip the memory encoder with `run_mem_encoder=False`. For example,
+        # in demo we might call `track_step` multiple times for each user click,
+        # and only encode the memory when the user finalizes their clicks. And in ablation
+        # settings like SAM training on static images, we don't need the memory encoder.
+        run_mem_encoder=True,
+        # The previously predicted SAM mask logits (which can be fed together with new clicks in demo).
+        prev_sam_mask_logits=None,
+    ):
+        current_out, sam_outputs, _, _ = self._track_step(
+            frame_idx,
+            is_init_cond_frame,
+            current_vision_feats,
+            current_vision_pos_embeds,
+            feat_sizes,
+            point_inputs,
+            mask_inputs,
+            output_dict,
+            num_frames,
+            track_in_reverse,
+            prev_sam_mask_logits,
+        )
+
+        (
+            _,
+            _,
+            _,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            object_score_logits,
+        ) = sam_outputs
+
+        current_out["pred_masks"] = low_res_masks
+        current_out["pred_masks_high_res"] = high_res_masks
+        current_out["obj_ptr"] = obj_ptr
+        if not self.training:
+            # Only add this in inference (to avoid unused param in activation checkpointing;
+            # it's mainly used in the demo to encode spatial memories w/ consolidated masks)
+            current_out["object_score_logits"] = object_score_logits
+
+        # Finally run the memory encoder on the predicted mask to encode
+        # it into a new memory feature (that can be used in future frames)
+        self._encode_memory_in_output(
+            current_vision_feats,
+            feat_sizes,
+            point_inputs,
+            run_mem_encoder,
+            high_res_masks,
+            object_score_logits,
+            current_out,
+        )
+
+        return current_out
+
+    def _use_multimask(self, is_init_cond_frame, point_inputs):
+        """Whether to use multimask output in the SAM head."""
+        num_pts = 0 if point_inputs is None else point_inputs["point_labels"].size(1)
+        multimask_output = (
+            self.multimask_output_in_sam
+            and (is_init_cond_frame or self.multimask_output_for_tracking)
+            and (self.multimask_min_pt_num <= num_pts <= self.multimask_max_pt_num)
+        )
+        return multimask_output
+
+    def _apply_non_overlapping_constraints(self, pred_masks):
+        """
+        Apply non-overlapping constraints to the object scores in pred_masks. Here we
+        keep only the highest scoring object at each spatial location in pred_masks.
+        """
+        batch_size = pred_masks.size(0)
+        if batch_size == 1:
+            return pred_masks
+
+        device = pred_masks.device
+        # "max_obj_inds": object index of the object with the highest score at each location
+        max_obj_inds = torch.argmax(pred_masks, dim=0, keepdim=True)
+        # "batch_obj_inds": object index of each object slice (along dim 0) in `pred_masks`
+        batch_obj_inds = torch.arange(batch_size, device=device)[:, None, None, None]
+        keep = max_obj_inds == batch_obj_inds
+        # suppress overlapping regions' scores below -10.0 so that the foreground regions
+        # don't overlap (here sigmoid(-10.0)=4.5398e-05)
+        pred_masks = torch.where(keep, pred_masks, torch.clamp(pred_masks, max=-10.0))
+        return pred_masks

custom_nodes/comfyui-segment-anything-2/sam2/modeling/sam2_utils.py

@@ -0,0 +1,323 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import copy
+from typing import Tuple
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..utils.misc import mask_to_box
+
+
+def select_closest_cond_frames(frame_idx, cond_frame_outputs, max_cond_frame_num):
+    """
+    Select up to `max_cond_frame_num` conditioning frames from `cond_frame_outputs`
+    that are temporally closest to the current frame at `frame_idx`. Here, we take
+    - a) the closest conditioning frame before `frame_idx` (if any);
+    - b) the closest conditioning frame after `frame_idx` (if any);
+    - c) any other temporally closest conditioning frames until reaching a total
+         of `max_cond_frame_num` conditioning frames.
+
+    Outputs:
+    - selected_outputs: selected items (keys & values) from `cond_frame_outputs`.
+    - unselected_outputs: items (keys & values) not selected in `cond_frame_outputs`.
+    """
+    if max_cond_frame_num == -1 or len(cond_frame_outputs) <= max_cond_frame_num:
+        selected_outputs = cond_frame_outputs
+        unselected_outputs = {}
+    else:
+        assert max_cond_frame_num >= 2, "we should allow using 2+ conditioning frames"
+        selected_outputs = {}
+
+        # the closest conditioning frame before `frame_idx` (if any)
+        idx_before = max((t for t in cond_frame_outputs if t < frame_idx), default=None)
+        if idx_before is not None:
+            selected_outputs[idx_before] = cond_frame_outputs[idx_before]
+
+        # the closest conditioning frame after `frame_idx` (if any)
+        idx_after = min((t for t in cond_frame_outputs if t >= frame_idx), default=None)
+        if idx_after is not None:
+            selected_outputs[idx_after] = cond_frame_outputs[idx_after]
+
+        # add other temporally closest conditioning frames until reaching a total
+        # of `max_cond_frame_num` conditioning frames.
+        num_remain = max_cond_frame_num - len(selected_outputs)
+        inds_remain = sorted(
+            (t for t in cond_frame_outputs if t not in selected_outputs),
+            key=lambda x: abs(x - frame_idx),
+        )[:num_remain]
+        selected_outputs.update((t, cond_frame_outputs[t]) for t in inds_remain)
+        unselected_outputs = {
+            t: v for t, v in cond_frame_outputs.items() if t not in selected_outputs
+        }
+
+    return selected_outputs, unselected_outputs
+
+
+def get_1d_sine_pe(pos_inds, dim, temperature=10000):
+    """
+    Get 1D sine positional embedding as in the original Transformer paper.
+    """
+    pe_dim = dim // 2
+    dim_t = torch.arange(pe_dim, dtype=torch.float32, device=pos_inds.device)
+    dim_t = temperature ** (2 * (dim_t // 2) / pe_dim)
+
+    pos_embed = pos_inds.unsqueeze(-1) / dim_t
+    pos_embed = torch.cat([pos_embed.sin(), pos_embed.cos()], dim=-1)
+    return pos_embed
+
+
+def get_activation_fn(activation):
+    """Return an activation function given a string"""
+    if activation == "relu":
+        return F.relu
+    if activation == "gelu":
+        return F.gelu
+    if activation == "glu":
+        return F.glu
+    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
+
+
+def get_clones(module, N):
+    return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
+
+
+class DropPath(nn.Module):
+    # adapted from https://github.com/huggingface/pytorch-image-models/blob/main/timm/layers/drop.py
+    def __init__(self, drop_prob=0.0, scale_by_keep=True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        if self.drop_prob == 0.0 or not self.training:
+            return x
+        keep_prob = 1 - self.drop_prob
+        shape = (x.shape[0],) + (1,) * (x.ndim - 1)
+        random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+        if keep_prob > 0.0 and self.scale_by_keep:
+            random_tensor.div_(keep_prob)
+        return x * random_tensor
+
+
+# Lightly adapted from
+# https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py # noqa
+class MLP(nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        output_dim: int,
+        num_layers: int,
+        activation: nn.Module = nn.ReLU,
+        sigmoid_output: bool = False,
+    ) -> None:
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(
+            nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])
+        )
+        self.sigmoid_output = sigmoid_output
+        self.act = activation()
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = self.act(layer(x)) if i < self.num_layers - 1 else layer(x)
+        if self.sigmoid_output:
+            x = F.sigmoid(x)
+        return x
+
+
+# From https://github.com/facebookresearch/detectron2/blob/main/detectron2/layers/batch_norm.py # noqa
+# Itself from https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119  # noqa
+class LayerNorm2d(nn.Module):
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x
+
+
+def sample_box_points(
+    masks: torch.Tensor,
+    noise: float = 0.1,  # SAM default
+    noise_bound: int = 20,  # SAM default
+    top_left_label: int = 2,
+    bottom_right_label: int = 3,
+) -> Tuple[np.array, np.array]:
+    """
+    Sample a noised version of the top left and bottom right corners of a given `bbox`
+
+    Inputs:
+    - masks: [B, 1, H,W] boxes, dtype=torch.Tensor
+    - noise: noise as a fraction of box width and height, dtype=float
+    - noise_bound: maximum amount of noise (in pure pixesl), dtype=int
+
+    Returns:
+    - box_coords: [B, num_pt, 2], contains (x, y) coordinates of top left and bottom right box corners, dtype=torch.float
+    - box_labels: [B, num_pt], label 2 is reserverd for top left and 3 for bottom right corners, dtype=torch.int32
+    """
+    device = masks.device
+    box_coords = mask_to_box(masks)
+    B, _, H, W = masks.shape
+    box_labels = torch.tensor(
+        [top_left_label, bottom_right_label], dtype=torch.int, device=device
+    ).repeat(B)
+    if noise > 0.0:
+        if not isinstance(noise_bound, torch.Tensor):
+            noise_bound = torch.tensor(noise_bound, device=device)
+        bbox_w = box_coords[..., 2] - box_coords[..., 0]
+        bbox_h = box_coords[..., 3] - box_coords[..., 1]
+        max_dx = torch.min(bbox_w * noise, noise_bound)
+        max_dy = torch.min(bbox_h * noise, noise_bound)
+        box_noise = 2 * torch.rand(B, 1, 4, device=device) - 1
+        box_noise = box_noise * torch.stack((max_dx, max_dy, max_dx, max_dy), dim=-1)
+
+        box_coords = box_coords + box_noise
+        img_bounds = (
+            torch.tensor([W, H, W, H], device=device) - 1
+        )  # uncentered pixel coords
+        box_coords.clamp_(torch.zeros_like(img_bounds), img_bounds)  # In place clamping
+
+    box_coords = box_coords.reshape(-1, 2, 2)  # always 2 points
+    box_labels = box_labels.reshape(-1, 2)
+    return box_coords, box_labels
+
+
+def sample_random_points_from_errors(gt_masks, pred_masks, num_pt=1):
+    """
+    Sample `num_pt` random points (along with their labels) independently from the error regions.
+
+    Inputs:
+    - gt_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool
+    - pred_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool or None
+    - num_pt: int, number of points to sample independently for each of the B error maps
+
+    Outputs:
+    - points: [B, num_pt, 2], dtype=torch.float, contains (x, y) coordinates of each sampled point
+    - labels: [B, num_pt], dtype=torch.int32, where 1 means positive clicks and 0 means
+      negative clicks
+    """
+    if pred_masks is None:  # if pred_masks is not provided, treat it as empty
+        pred_masks = torch.zeros_like(gt_masks)
+    assert gt_masks.dtype == torch.bool and gt_masks.size(1) == 1
+    assert pred_masks.dtype == torch.bool and pred_masks.shape == gt_masks.shape
+    assert num_pt >= 0
+
+    B, _, H_im, W_im = gt_masks.shape
+    device = gt_masks.device
+
+    # false positive region, a new point sampled in this region should have
+    # negative label to correct the FP error
+    fp_masks = ~gt_masks & pred_masks
+    # false negative region, a new point sampled in this region should have
+    # positive label to correct the FN error
+    fn_masks = gt_masks & ~pred_masks
+    # whether the prediction completely match the ground-truth on each mask
+    all_correct = torch.all((gt_masks == pred_masks).flatten(2), dim=2)
+    all_correct = all_correct[..., None, None]
+
+    # channel 0 is FP map, while channel 1 is FN map
+    pts_noise = torch.rand(B, num_pt, H_im, W_im, 2, device=device)
+    # sample a negative new click from FP region or a positive new click
+    # from FN region, depend on where the maximum falls,
+    # and in case the predictions are all correct (no FP or FN), we just
+    # sample a negative click from the background region
+    pts_noise[..., 0] *= fp_masks | (all_correct & ~gt_masks)
+    pts_noise[..., 1] *= fn_masks
+    pts_idx = pts_noise.flatten(2).argmax(dim=2)
+    labels = (pts_idx % 2).to(torch.int32)
+    pts_idx = pts_idx // 2
+    pts_x = pts_idx % W_im
+    pts_y = pts_idx // W_im
+    points = torch.stack([pts_x, pts_y], dim=2).to(torch.float)
+    return points, labels
+
+
+def sample_one_point_from_error_center(gt_masks, pred_masks, padding=True):
+    """
+    Sample 1 random point (along with its label) from the center of each error region,
+    that is, the point with the largest distance to the boundary of each error region.
+    This is the RITM sampling method from https://github.com/saic-vul/ritm_interactive_segmentation/blob/master/isegm/inference/clicker.py
+
+    Inputs:
+    - gt_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool
+    - pred_masks: [B, 1, H_im, W_im] masks, dtype=torch.bool or None
+    - padding: if True, pad with boundary of 1 px for distance transform
+
+    Outputs:
+    - points: [B, 1, 2], dtype=torch.float, contains (x, y) coordinates of each sampled point
+    - labels: [B, 1], dtype=torch.int32, where 1 means positive clicks and 0 means negative clicks
+    """
+    import cv2
+
+    if pred_masks is None:
+        pred_masks = torch.zeros_like(gt_masks)
+    assert gt_masks.dtype == torch.bool and gt_masks.size(1) == 1
+    assert pred_masks.dtype == torch.bool and pred_masks.shape == gt_masks.shape
+
+    B, _, _, W_im = gt_masks.shape
+    device = gt_masks.device
+
+    # false positive region, a new point sampled in this region should have
+    # negative label to correct the FP error
+    fp_masks = ~gt_masks & pred_masks
+    # false negative region, a new point sampled in this region should have
+    # positive label to correct the FN error
+    fn_masks = gt_masks & ~pred_masks
+
+    fp_masks = fp_masks.cpu().numpy()
+    fn_masks = fn_masks.cpu().numpy()
+    points = torch.zeros(B, 1, 2, dtype=torch.float)
+    labels = torch.ones(B, 1, dtype=torch.int32)
+    for b in range(B):
+        fn_mask = fn_masks[b, 0]
+        fp_mask = fp_masks[b, 0]
+        if padding:
+            fn_mask = np.pad(fn_mask, ((1, 1), (1, 1)), "constant")
+            fp_mask = np.pad(fp_mask, ((1, 1), (1, 1)), "constant")
+        # compute the distance of each point in FN/FP region to its boundary
+        fn_mask_dt = cv2.distanceTransform(fn_mask.astype(np.uint8), cv2.DIST_L2, 0)
+        fp_mask_dt = cv2.distanceTransform(fp_mask.astype(np.uint8), cv2.DIST_L2, 0)
+        if padding:
+            fn_mask_dt = fn_mask_dt[1:-1, 1:-1]
+            fp_mask_dt = fp_mask_dt[1:-1, 1:-1]
+
+        # take the point in FN/FP region with the largest distance to its boundary
+        fn_mask_dt_flat = fn_mask_dt.reshape(-1)
+        fp_mask_dt_flat = fp_mask_dt.reshape(-1)
+        fn_argmax = np.argmax(fn_mask_dt_flat)
+        fp_argmax = np.argmax(fp_mask_dt_flat)
+        is_positive = fn_mask_dt_flat[fn_argmax] > fp_mask_dt_flat[fp_argmax]
+        pt_idx = fn_argmax if is_positive else fp_argmax
+        points[b, 0, 0] = pt_idx % W_im  # x
+        points[b, 0, 1] = pt_idx // W_im  # y
+        labels[b, 0] = int(is_positive)
+
+    points = points.to(device)
+    labels = labels.to(device)
+    return points, labels
+
+
+def get_next_point(gt_masks, pred_masks, method):
+    if method == "uniform":
+        return sample_random_points_from_errors(gt_masks, pred_masks)
+    elif method == "center":
+        return sample_one_point_from_error_center(gt_masks, pred_masks)
+    else:
+        raise ValueError(f"unknown sampling method {method}")

custom_nodes/comfyui-segment-anything-2/sam2/sam2_image_predictor.py

@@ -0,0 +1,446 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from PIL.Image import Image
+
+from ..sam2.modeling.sam2_base import SAM2Base
+
+from ..sam2.utils.transforms import SAM2Transforms
+
+
+class SAM2ImagePredictor:
+    def __init__(
+        self,
+        sam_model: SAM2Base,
+        mask_threshold=0.0,
+        max_hole_area=0.0,
+        max_sprinkle_area=0.0,
+    ) -> None:
+        """
+        Uses SAM-2 to calculate the image embedding for an image, and then
+        allow repeated, efficient mask prediction given prompts.
+
+        Arguments:
+          sam_model (Sam-2): The model to use for mask prediction.
+          mask_threshold (float): The threshold to use when converting mask logits
+            to binary masks. Masks are thresholded at 0 by default.
+          fill_hole_area (int): If fill_hole_area > 0, we fill small holes in up to
+            the maximum area of fill_hole_area in low_res_masks.
+        """
+        super().__init__()
+        self.model = sam_model
+        self._transforms = SAM2Transforms(
+            resolution=self.model.image_size,
+            mask_threshold=mask_threshold,
+            max_hole_area=max_hole_area,
+            max_sprinkle_area=max_sprinkle_area,
+        )
+
+        # Predictor state
+        self._is_image_set = False
+        self._features = None
+        self._orig_hw = None
+        # Whether the predictor is set for single image or a batch of images
+        self._is_batch = False
+
+        # Predictor config
+        self.mask_threshold = mask_threshold
+
+        # Spatial dim for backbone feature maps
+        self._bb_feat_sizes = [
+            (256, 256),
+            (128, 128),
+            (64, 64),
+        ]
+
+    @torch.no_grad()
+    def set_image(
+        self,
+        image: Union[np.ndarray, Image],
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image, allowing
+        masks to be predicted with the 'predict' method.
+
+        Arguments:
+          image (np.ndarray or PIL Image): The input image to embed in RGB format. The image should be in HWC format if np.ndarray, or WHC format if PIL Image
+          with pixel values in [0, 255].
+          image_format (str): The color format of the image, in ['RGB', 'BGR'].
+        """
+        self.reset_predictor()
+        # Transform the image to the form expected by the model
+        if isinstance(image, np.ndarray):
+            #logging.info("For numpy array image, we assume (HxWxC) format")
+            self._orig_hw = [image.shape[:2]]
+        elif isinstance(image, Image):
+            w, h = image.size
+            self._orig_hw = [(h, w)]
+        else:
+            raise NotImplementedError("Image format not supported")
+
+        input_image = self._transforms(image)
+        input_image = input_image[None, ...].to(self.device)
+
+        assert (
+            len(input_image.shape) == 4 and input_image.shape[1] == 3
+        ), f"input_image must be of size 1x3xHxW, got {input_image.shape}"
+        #logging.info("Computing image embeddings for the provided image...")
+        backbone_out = self.model.forward_image(input_image)
+        _, vision_feats, _, _ = self.model._prepare_backbone_features(backbone_out)
+        # Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
+        if self.model.directly_add_no_mem_embed:
+            vision_feats[-1] = vision_feats[-1] + self.model.no_mem_embed
+
+        feats = [
+            feat.permute(1, 2, 0).view(1, -1, *feat_size)
+            for feat, feat_size in zip(vision_feats[::-1], self._bb_feat_sizes[::-1])
+        ][::-1]
+        self._features = {"image_embed": feats[-1], "high_res_feats": feats[:-1]}
+        self._is_image_set = True
+        #logging.info("Image embeddings computed.")
+
+    @torch.no_grad()
+    def set_image_batch(
+        self,
+        image_list: List[Union[np.ndarray]],
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image batch, allowing
+        masks to be predicted with the 'predict_batch' method.
+
+        Arguments:
+          image_list (List[np.ndarray]): The input images to embed in RGB format. The image should be in HWC format if np.ndarray
+          with pixel values in [0, 255].
+        """
+        self.reset_predictor()
+        assert isinstance(image_list, list)
+        self._orig_hw = []
+        for image in image_list:
+            assert isinstance(
+                image, np.ndarray
+            ), "Images are expected to be an np.ndarray in RGB format, and of shape  HWC"
+            self._orig_hw.append(image.shape[:2])
+        # Transform the image to the form expected by the model
+        img_batch = self._transforms.forward_batch(image_list)
+        img_batch = img_batch.to(self.device)
+        batch_size = img_batch.shape[0]
+        assert (
+            len(img_batch.shape) == 4 and img_batch.shape[1] == 3
+        ), f"img_batch must be of size Bx3xHxW, got {img_batch.shape}"
+        logging.info("Computing image embeddings for the provided images...")
+        backbone_out = self.model.forward_image(img_batch)
+        _, vision_feats, _, _ = self.model._prepare_backbone_features(backbone_out)
+        # Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
+        if self.model.directly_add_no_mem_embed:
+            vision_feats[-1] = vision_feats[-1] + self.model.no_mem_embed
+
+        feats = [
+            feat.permute(1, 2, 0).view(batch_size, -1, *feat_size)
+            for feat, feat_size in zip(vision_feats[::-1], self._bb_feat_sizes[::-1])
+        ][::-1]
+        self._features = {"image_embed": feats[-1], "high_res_feats": feats[:-1]}
+        self._is_image_set = True
+        self._is_batch = True
+        logging.info("Image embeddings computed.")
+
+    def predict_batch(
+        self,
+        point_coords_batch: List[np.ndarray] = None,
+        point_labels_batch: List[np.ndarray] = None,
+        box_batch: List[np.ndarray] = None,
+        mask_input_batch: List[np.ndarray] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        normalize_coords=True,
+    ) -> Tuple[List[np.ndarray], List[np.ndarray], List[np.ndarray]]:
+        """This function is very similar to predict(...), however it is used for batched mode, when the model is expected to generate predictions on multiple images.
+        It returns a tupele of lists of masks, ious, and low_res_masks_logits.
+        """
+        assert self._is_batch, "This function should only be used when in batched mode"
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image_batch(...) before mask prediction."
+            )
+        num_images = len(self._features["image_embed"])
+        all_masks = []
+        all_ious = []
+        all_low_res_masks = []
+        for img_idx in range(num_images):
+            # Transform input prompts
+            point_coords = (
+                point_coords_batch[img_idx] if point_coords_batch is not None else None
+            )
+            point_labels = (
+                point_labels_batch[img_idx] if point_labels_batch is not None else None
+            )
+            box = box_batch[img_idx] if box_batch is not None else None
+            mask_input = (
+                mask_input_batch[img_idx] if mask_input_batch is not None else None
+            )
+            mask_input, unnorm_coords, labels, unnorm_box = self._prep_prompts(
+                point_coords,
+                point_labels,
+                box,
+                mask_input,
+                normalize_coords,
+                img_idx=img_idx,
+            )
+            masks, iou_predictions, low_res_masks = self._predict(
+                unnorm_coords,
+                labels,
+                unnorm_box,
+                mask_input,
+                multimask_output,
+                return_logits=return_logits,
+                img_idx=img_idx,
+            )
+            masks_np = masks.squeeze(0).float().detach().cpu().numpy()
+            iou_predictions_np = (
+                iou_predictions.squeeze(0).float().detach().cpu().numpy()
+            )
+            low_res_masks_np = low_res_masks.squeeze(0).float().detach().cpu().numpy()
+            all_masks.append(masks_np)
+            all_ious.append(iou_predictions_np)
+            all_low_res_masks.append(low_res_masks_np)
+
+        return all_masks, all_ious, all_low_res_masks
+
+    def predict(
+        self,
+        point_coords: Optional[np.ndarray] = None,
+        point_labels: Optional[np.ndarray] = None,
+        box: Optional[np.ndarray] = None,
+        mask_input: Optional[np.ndarray] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        normalize_coords=True,
+    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+
+        Arguments:
+          point_coords (np.ndarray or None): A Nx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (np.ndarray or None): A length N array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          box (np.ndarray or None): A length 4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form 1xHxW, where
+            for SAM, H=W=256.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+          normalize_coords (bool): If true, the point coordinates will be normalized to the range [0,1] and point_coords is expected to be wrt. image dimensions.
+
+        Returns:
+          (np.ndarray): The output masks in CxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (np.ndarray): An array of length C containing the model's
+            predictions for the quality of each mask.
+          (np.ndarray): An array of shape CxHxW, where C is the number
+            of masks and H=W=256. These low resolution logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) before mask prediction."
+            )
+
+        # Transform input prompts
+
+        mask_input, unnorm_coords, labels, unnorm_box = self._prep_prompts(
+            point_coords, point_labels, box, mask_input, normalize_coords
+        )
+
+        masks, iou_predictions, low_res_masks = self._predict(
+            unnorm_coords,
+            labels,
+            unnorm_box,
+            mask_input,
+            multimask_output,
+            return_logits=return_logits,
+        )
+
+        masks_np = masks.squeeze(0).float().detach().cpu().numpy()
+        iou_predictions_np = iou_predictions.squeeze(0).float().detach().cpu().numpy()
+        low_res_masks_np = low_res_masks.squeeze(0).float().detach().cpu().numpy()
+        return masks_np, iou_predictions_np, low_res_masks_np
+
+    def _prep_prompts(
+        self, point_coords, point_labels, box, mask_logits, normalize_coords, img_idx=-1
+    ):
+
+        unnorm_coords, labels, unnorm_box, mask_input = None, None, None, None
+        if point_coords is not None:
+            assert (
+                point_labels is not None
+            ), "point_labels must be supplied if point_coords is supplied."
+            point_coords = torch.as_tensor(
+                point_coords, dtype=torch.float, device=self.device
+            )
+            unnorm_coords = self._transforms.transform_coords(
+                point_coords, normalize=normalize_coords, orig_hw=self._orig_hw[img_idx]
+            )
+            labels = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)
+            if len(unnorm_coords.shape) == 2:
+                unnorm_coords, labels = unnorm_coords[None, ...], labels[None, ...]
+        if box is not None:
+            box = torch.as_tensor(box, dtype=torch.float, device=self.device)
+            unnorm_box = self._transforms.transform_boxes(
+                box, normalize=normalize_coords, orig_hw=self._orig_hw[img_idx]
+            )  # Bx2x2
+        if mask_logits is not None:
+            mask_input = torch.as_tensor(
+                mask_logits, dtype=torch.float, device=self.device
+            )
+            if len(mask_input.shape) == 3:
+                mask_input = mask_input[None, :, :, :]
+        return mask_input, unnorm_coords, labels, unnorm_box
+
+    @torch.no_grad()
+    def _predict(
+        self,
+        point_coords: Optional[torch.Tensor],
+        point_labels: Optional[torch.Tensor],
+        boxes: Optional[torch.Tensor] = None,
+        mask_input: Optional[torch.Tensor] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        img_idx: int = -1,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+        Input prompts are batched torch tensors and are expected to already be
+        transformed to the input frame using SAM2Transforms.
+
+        Arguments:
+          point_coords (torch.Tensor or None): A BxNx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (torch.Tensor or None): A BxN array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          boxes (np.ndarray or None): A Bx4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form Bx1xHxW, where
+            for SAM, H=W=256. Masks returned by a previous iteration of the
+            predict method do not need further transformation.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+
+        Returns:
+          (torch.Tensor): The output masks in BxCxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (torch.Tensor): An array of shape BxC containing the model's
+            predictions for the quality of each mask.
+          (torch.Tensor): An array of shape BxCxHxW, where C is the number
+            of masks and H=W=256. These low res logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) before mask prediction."
+            )
+
+        if point_coords is not None:
+            concat_points = (point_coords, point_labels)
+        else:
+            concat_points = None
+
+        # Embed prompts
+        if boxes is not None:
+            box_coords = boxes.reshape(-1, 2, 2)
+            box_labels = torch.tensor([[2, 3]], dtype=torch.int, device=boxes.device)
+            box_labels = box_labels.repeat(boxes.size(0), 1)
+            # we merge "boxes" and "points" into a single "concat_points" input (where
+            # boxes are added at the beginning) to sam_prompt_encoder
+            if concat_points is not None:
+                concat_coords = torch.cat([box_coords, concat_points[0]], dim=1)
+                concat_labels = torch.cat([box_labels, concat_points[1]], dim=1)
+                concat_points = (concat_coords, concat_labels)
+            else:
+                concat_points = (box_coords, box_labels)
+
+        sparse_embeddings, dense_embeddings = self.model.sam_prompt_encoder(
+            points=concat_points,
+            boxes=None,
+            masks=mask_input,
+        )
+
+        # Predict masks
+        batched_mode = (
+            concat_points is not None and concat_points[0].shape[0] > 1
+        )  # multi object prediction
+        high_res_features = [
+            feat_level[img_idx].unsqueeze(0)
+            for feat_level in self._features["high_res_feats"]
+        ]
+        low_res_masks, iou_predictions, _, _ = self.model.sam_mask_decoder(
+            image_embeddings=self._features["image_embed"][img_idx].unsqueeze(0),
+            image_pe=self.model.sam_prompt_encoder.get_dense_pe(),
+            sparse_prompt_embeddings=sparse_embeddings,
+            dense_prompt_embeddings=dense_embeddings,
+            multimask_output=multimask_output,
+            repeat_image=batched_mode,
+            high_res_features=high_res_features,
+        )
+
+        # Upscale the masks to the original image resolution
+        masks = self._transforms.postprocess_masks(
+            low_res_masks, self._orig_hw[img_idx]
+        )
+        low_res_masks = torch.clamp(low_res_masks, -32.0, 32.0)
+        if not return_logits:
+            masks = masks > self.mask_threshold
+
+        return masks, iou_predictions, low_res_masks
+
+    def get_image_embedding(self) -> torch.Tensor:
+        """
+        Returns the image embeddings for the currently set image, with
+        shape 1xCxHxW, where C is the embedding dimension and (H,W) are
+        the embedding spatial dimension of SAM (typically C=256, H=W=64).
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) to generate an embedding."
+            )
+        assert (
+            self._features is not None
+        ), "Features must exist if an image has been set."
+        return self._features["image_embed"]
+
+    @property
+    def device(self) -> torch.device:
+        return self.model.device
+
+    def reset_predictor(self) -> None:
+        """
+        Resets the image embeddings and other state variables.
+        """
+        self._is_image_set = False
+        self._features = None
+        self._orig_hw = None
+        self._is_batch = False

custom_nodes/comfyui-segment-anything-2/sam2/sam2_video_predictor.py

@@ -0,0 +1,1154 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import warnings
+from collections import OrderedDict
+
+import torch
+
+from tqdm import tqdm
+
+from ..sam2.modeling.sam2_base import NO_OBJ_SCORE, SAM2Base
+from ..sam2.utils.misc import concat_points, fill_holes_in_mask_scores, load_video_frames
+
+
+class SAM2VideoPredictor(SAM2Base):
+    """The predictor class to handle user interactions and manage inference states."""
+
+    def __init__(
+        self,
+        fill_hole_area=0,
+        # whether to apply non-overlapping constraints on the output object masks
+        non_overlap_masks=False,
+        # whether to clear non-conditioning memory of the surrounding frames (which may contain outdated information) after adding correction clicks;
+        # note that this would only apply to *single-object tracking* unless `clear_non_cond_mem_for_multi_obj` is also set to True)
+        clear_non_cond_mem_around_input=False,
+        # whether to also clear non-conditioning memory of the surrounding frames (only effective when `clear_non_cond_mem_around_input` is True).
+        clear_non_cond_mem_for_multi_obj=False,
+        # if `add_all_frames_to_correct_as_cond` is True, we also append to the conditioning frame list any frame that receives a later correction click
+        # if `add_all_frames_to_correct_as_cond` is False, we conditioning frame list to only use those initial conditioning frames
+        add_all_frames_to_correct_as_cond=False,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.fill_hole_area = fill_hole_area
+        self.non_overlap_masks = non_overlap_masks
+        self.clear_non_cond_mem_around_input = clear_non_cond_mem_around_input
+        self.clear_non_cond_mem_for_multi_obj = clear_non_cond_mem_for_multi_obj
+        self.add_all_frames_to_correct_as_cond = add_all_frames_to_correct_as_cond
+
+    @torch.inference_mode()
+    def init_state(
+        self,
+        images,
+        video_height,
+        video_width,
+        device='cuda',
+        offload_video_to_cpu=False,
+        offload_state_to_cpu=False,
+        async_loading_frames=False,
+    ):
+        """Initialize a inference state."""
+        # images, video_height, video_width = load_video_frames(
+        #     video_path=video_path,
+        #     image_size=self.image_size,
+        #     offload_video_to_cpu=offload_video_to_cpu,
+        #     async_loading_frames=async_loading_frames,
+        # )
+        inference_state = {}
+        inference_state["images"] = images
+        inference_state["num_frames"] = len(images)
+        # whether to offload the video frames to CPU memory
+        # turning on this option saves the GPU memory with only a very small overhead
+        inference_state["offload_video_to_cpu"] = offload_video_to_cpu
+        # whether to offload the inference state to CPU memory
+        # turning on this option saves the GPU memory at the cost of a lower tracking fps
+        # (e.g. in a test case of 768x768 model, fps dropped from 27 to 24 when tracking one object
+        # and from 24 to 21 when tracking two objects)
+        inference_state["offload_state_to_cpu"] = offload_state_to_cpu
+        # the original video height and width, used for resizing final output scores
+        inference_state["video_height"] = video_height
+        inference_state["video_width"] = video_width
+        inference_state["device"] = torch.device(device)
+        if offload_state_to_cpu:
+            inference_state["storage_device"] = torch.device("cpu")
+        else:
+            inference_state["storage_device"] = torch.device(device)
+        # inputs on each frame
+        inference_state["point_inputs_per_obj"] = {}
+        inference_state["mask_inputs_per_obj"] = {}
+        # visual features on a small number of recently visited frames for quick interactions
+        inference_state["cached_features"] = {}
+        # values that don't change across frames (so we only need to hold one copy of them)
+        inference_state["constants"] = {}
+        # mapping between client-side object id and model-side object index
+        inference_state["obj_id_to_idx"] = OrderedDict()
+        inference_state["obj_idx_to_id"] = OrderedDict()
+        inference_state["obj_ids"] = []
+        # A storage to hold the model's tracking results and states on each frame
+        inference_state["output_dict"] = {
+            "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+        }
+        # Slice (view) of each object tracking results, sharing the same memory with "output_dict"
+        inference_state["output_dict_per_obj"] = {}
+        # A temporary storage to hold new outputs when user interact with a frame
+        # to add clicks or mask (it's merged into "output_dict" before propagation starts)
+        inference_state["temp_output_dict_per_obj"] = {}
+        # Frames that already holds consolidated outputs from click or mask inputs
+        # (we directly use their consolidated outputs during tracking)
+        inference_state["consolidated_frame_inds"] = {
+            "cond_frame_outputs": set(),  # set containing frame indices
+            "non_cond_frame_outputs": set(),  # set containing frame indices
+        }
+        # metadata for each tracking frame (e.g. which direction it's tracked)
+        inference_state["tracking_has_started"] = False
+        inference_state["frames_already_tracked"] = {}
+        # Warm up the visual backbone and cache the image feature on frame 0
+        self._get_image_feature(inference_state, frame_idx=0, batch_size=1)
+        return inference_state
+
+    def _obj_id_to_idx(self, inference_state, obj_id):
+        """Map client-side object id to model-side object index."""
+        obj_idx = inference_state["obj_id_to_idx"].get(obj_id, None)
+        if obj_idx is not None:
+            return obj_idx
+
+        # This is a new object id not sent to the server before. We only allow adding
+        # new objects *before* the tracking starts.
+        allow_new_object = not inference_state["tracking_has_started"]
+        if allow_new_object:
+            # get the next object slot
+            obj_idx = len(inference_state["obj_id_to_idx"])
+            inference_state["obj_id_to_idx"][obj_id] = obj_idx
+            inference_state["obj_idx_to_id"][obj_idx] = obj_id
+            inference_state["obj_ids"] = list(inference_state["obj_id_to_idx"])
+            # set up input and output structures for this object
+            inference_state["point_inputs_per_obj"][obj_idx] = {}
+            inference_state["mask_inputs_per_obj"][obj_idx] = {}
+            inference_state["output_dict_per_obj"][obj_idx] = {
+                "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+                "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            }
+            inference_state["temp_output_dict_per_obj"][obj_idx] = {
+                "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+                "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            }
+            return obj_idx
+        else:
+            raise RuntimeError(
+                f"Cannot add new object id {obj_id} after tracking starts. "
+                f"All existing object ids: {inference_state['obj_ids']}. "
+                f"Please call 'reset_state' to restart from scratch."
+            )
+
+    def _obj_idx_to_id(self, inference_state, obj_idx):
+        """Map model-side object index to client-side object id."""
+        return inference_state["obj_idx_to_id"][obj_idx]
+
+    def _get_obj_num(self, inference_state):
+        """Get the total number of unique object ids received so far in this session."""
+        return len(inference_state["obj_idx_to_id"])
+
+    @torch.inference_mode()
+    def add_new_points_or_box(
+        self,
+        inference_state,
+        frame_idx,
+        obj_id,
+        points=None,
+        labels=None,
+        clear_old_points=True,
+        normalize_coords=True,
+        box=None,
+    ):
+        """Add new points to a frame."""
+        obj_idx = self._obj_id_to_idx(inference_state, obj_id)
+        point_inputs_per_frame = inference_state["point_inputs_per_obj"][obj_idx]
+        mask_inputs_per_frame = inference_state["mask_inputs_per_obj"][obj_idx]
+
+        if (points is not None) != (labels is not None):
+            raise ValueError("points and labels must be provided together")
+        if points is None and box is None:
+            raise ValueError("at least one of points or box must be provided as input")
+
+        if points is None:
+            points = torch.zeros(0, 2, dtype=torch.float32)
+        elif not isinstance(points, torch.Tensor):
+            points = torch.tensor(points, dtype=torch.float32)
+        if labels is None:
+            labels = torch.zeros(0, dtype=torch.int32)
+        elif not isinstance(labels, torch.Tensor):
+            labels = torch.tensor(labels, dtype=torch.int32)
+        if points.dim() == 2:
+            points = points.unsqueeze(0)  # add batch dimension
+        if labels.dim() == 1:
+            labels = labels.unsqueeze(0)  # add batch dimension
+
+        # If `box` is provided, we add it as the first two points with labels 2 and 3
+        # along with the user-provided points (consistent with how SAM 2 is trained).
+        if box is not None:
+            if not clear_old_points:
+                raise ValueError(
+                    "cannot add box without clearing old points, since "
+                    "box prompt must be provided before any point prompt "
+                    "(please use clear_old_points=True instead)"
+                )
+            if inference_state["tracking_has_started"]:
+                warnings.warn(
+                    "You are adding a box after tracking starts. SAM 2 may not always be "
+                    "able to incorporate a box prompt for *refinement*. If you intend to "
+                    "use box prompt as an *initial* input before tracking, please call "
+                    "'reset_state' on the inference state to restart from scratch.",
+                    category=UserWarning,
+                    stacklevel=2,
+                )
+            if not isinstance(box, torch.Tensor):
+                box = torch.tensor(box, dtype=torch.float32, device=points.device)
+            box_coords = box.reshape(1, 2, 2)
+            box_labels = torch.tensor([2, 3], dtype=torch.int32, device=labels.device)
+            box_labels = box_labels.reshape(1, 2)
+            points = torch.cat([box_coords, points], dim=1)
+            labels = torch.cat([box_labels, labels], dim=1)
+
+        if normalize_coords:
+            video_H = inference_state["video_height"]
+            video_W = inference_state["video_width"]
+            points = points / torch.tensor([video_W, video_H]).to(points.device)
+        # scale the (normalized) coordinates by the model's internal image size
+        points = points * self.image_size
+        points = points.to(inference_state["device"])
+        labels = labels.to(inference_state["device"])
+
+        if not clear_old_points:
+            point_inputs = point_inputs_per_frame.get(frame_idx, None)
+        else:
+            point_inputs = None
+        point_inputs = concat_points(point_inputs, points, labels)
+
+        point_inputs_per_frame[frame_idx] = point_inputs
+        mask_inputs_per_frame.pop(frame_idx, None)
+        # If this frame hasn't been tracked before, we treat it as an initial conditioning
+        # frame, meaning that the inputs points are to generate segments on this frame without
+        # using any memory from other frames, like in SAM. Otherwise (if it has been tracked),
+        # the input points will be used to correct the already tracked masks.
+        is_init_cond_frame = frame_idx not in inference_state["frames_already_tracked"]
+        # whether to track in reverse time order
+        if is_init_cond_frame:
+            reverse = False
+        else:
+            reverse = inference_state["frames_already_tracked"][frame_idx]["reverse"]
+        obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+        obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+        # Add a frame to conditioning output if it's an initial conditioning frame or
+        # if the model sees all frames receiving clicks/mask as conditioning frames.
+        is_cond = is_init_cond_frame or self.add_all_frames_to_correct_as_cond
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+
+        # Get any previously predicted mask logits on this object and feed it along with
+        # the new clicks into the SAM mask decoder.
+        prev_sam_mask_logits = None
+        # lookup temporary output dict first, which contains the most recent output
+        # (if not found, then lookup conditioning and non-conditioning frame output)
+        prev_out = obj_temp_output_dict[storage_key].get(frame_idx)
+        if prev_out is None:
+            prev_out = obj_output_dict["cond_frame_outputs"].get(frame_idx)
+            if prev_out is None:
+                prev_out = obj_output_dict["non_cond_frame_outputs"].get(frame_idx)
+
+        if prev_out is not None and prev_out["pred_masks"] is not None:
+            prev_sam_mask_logits = prev_out["pred_masks"].to(inference_state["device"],non_blocking=True)
+            # Clamp the scale of prev_sam_mask_logits to avoid rare numerical issues.
+            prev_sam_mask_logits = torch.clamp(prev_sam_mask_logits, -32.0, 32.0)
+        current_out, _ = self._run_single_frame_inference(
+            inference_state=inference_state,
+            output_dict=obj_output_dict,  # run on the slice of a single object
+            frame_idx=frame_idx,
+            batch_size=1,  # run on the slice of a single object
+            is_init_cond_frame=is_init_cond_frame,
+            point_inputs=point_inputs,
+            mask_inputs=None,
+            reverse=reverse,
+            # Skip the memory encoder when adding clicks or mask. We execute the memory encoder
+            # at the beginning of `propagate_in_video` (after user finalize their clicks). This
+            # allows us to enforce non-overlapping constraints on all objects before encoding
+            # them into memory.
+            run_mem_encoder=False,
+            prev_sam_mask_logits=prev_sam_mask_logits,
+        )
+        # Add the output to the output dict (to be used as future memory)
+        obj_temp_output_dict[storage_key][frame_idx] = current_out
+
+        # Resize the output mask to the original video resolution
+        obj_ids = inference_state["obj_ids"]
+        consolidated_out = self._consolidate_temp_output_across_obj(
+            inference_state,
+            frame_idx,
+            is_cond=is_cond,
+            run_mem_encoder=False,
+            consolidate_at_video_res=True,
+        )
+        _, video_res_masks = self._get_orig_video_res_output(
+            inference_state, consolidated_out["pred_masks_video_res"]
+        )
+        return frame_idx, obj_ids, video_res_masks
+
+    def add_new_points(self, *args, **kwargs):
+        """Deprecated method. Please use `add_new_points_or_box` instead."""
+        return self.add_new_points_or_box(*args, **kwargs)
+
+    @torch.inference_mode()
+    def add_new_mask(
+        self,
+        inference_state,
+        frame_idx,
+        obj_id,
+        mask,
+    ):
+        """Add new mask to a frame."""
+        obj_idx = self._obj_id_to_idx(inference_state, obj_id)
+        point_inputs_per_frame = inference_state["point_inputs_per_obj"][obj_idx]
+        mask_inputs_per_frame = inference_state["mask_inputs_per_obj"][obj_idx]
+
+        if not isinstance(mask, torch.Tensor):
+            mask = torch.tensor(mask, dtype=torch.bool)
+        assert mask.dim() == 2
+        mask_H, mask_W = mask.shape
+        mask_inputs_orig = mask[None, None]  # add batch and channel dimension
+        mask_inputs_orig = mask_inputs_orig.float().to(inference_state["device"])
+
+        # resize the mask if it doesn't match the model's image size
+        if mask_H != self.image_size or mask_W != self.image_size:
+            mask_inputs = torch.nn.functional.interpolate(
+                mask_inputs_orig,
+                size=(self.image_size, self.image_size),
+                align_corners=False,
+                mode="bilinear",
+                antialias=True,  # use antialias for downsampling
+            )
+            mask_inputs = (mask_inputs >= 0.5).float()
+        else:
+            mask_inputs = mask_inputs_orig
+
+        mask_inputs_per_frame[frame_idx] = mask_inputs
+        point_inputs_per_frame.pop(frame_idx, None)
+        # If this frame hasn't been tracked before, we treat it as an initial conditioning
+        # frame, meaning that the inputs points are to generate segments on this frame without
+        # using any memory from other frames, like in SAM. Otherwise (if it has been tracked),
+        # the input points will be used to correct the already tracked masks.
+        is_init_cond_frame = frame_idx not in inference_state["frames_already_tracked"]
+        # whether to track in reverse time order
+        if is_init_cond_frame:
+            reverse = False
+        else:
+            reverse = inference_state["frames_already_tracked"][frame_idx]["reverse"]
+        obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+        obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+        # Add a frame to conditioning output if it's an initial conditioning frame or
+        # if the model sees all frames receiving clicks/mask as conditioning frames.
+        is_cond = is_init_cond_frame or self.add_all_frames_to_correct_as_cond
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+
+        current_out, _ = self._run_single_frame_inference(
+            inference_state=inference_state,
+            output_dict=obj_output_dict,  # run on the slice of a single object
+            frame_idx=frame_idx,
+            batch_size=1,  # run on the slice of a single object
+            is_init_cond_frame=is_init_cond_frame,
+            point_inputs=None,
+            mask_inputs=mask_inputs,
+            reverse=reverse,
+            # Skip the memory encoder when adding clicks or mask. We execute the memory encoder
+            # at the beginning of `propagate_in_video` (after user finalize their clicks). This
+            # allows us to enforce non-overlapping constraints on all objects before encoding
+            # them into memory.
+            run_mem_encoder=False,
+        )
+        # Add the output to the output dict (to be used as future memory)
+        obj_temp_output_dict[storage_key][frame_idx] = current_out
+
+        # Resize the output mask to the original video resolution
+        obj_ids = inference_state["obj_ids"]
+        consolidated_out = self._consolidate_temp_output_across_obj(
+            inference_state,
+            frame_idx,
+            is_cond=is_cond,
+            run_mem_encoder=False,
+            consolidate_at_video_res=True,
+        )
+        _, video_res_masks = self._get_orig_video_res_output(
+            inference_state, consolidated_out["pred_masks_video_res"]
+        )
+        return frame_idx, obj_ids, video_res_masks
+
+    def _get_orig_video_res_output(self, inference_state, any_res_masks):
+        """
+        Resize the object scores to the original video resolution (video_res_masks)
+        and apply non-overlapping constraints for final output.
+        """
+        device = inference_state["device"]
+        video_H = inference_state["video_height"]
+        video_W = inference_state["video_width"]
+        any_res_masks = any_res_masks.to(device, non_blocking=True)
+        if any_res_masks.shape[-2:] == (video_H, video_W):
+            video_res_masks = any_res_masks
+        else:
+            video_res_masks = torch.nn.functional.interpolate(
+                any_res_masks,
+                size=(video_H, video_W),
+                mode="bilinear",
+                align_corners=False,
+            )
+        if self.non_overlap_masks:
+            video_res_masks = self._apply_non_overlapping_constraints(video_res_masks)
+        return any_res_masks, video_res_masks
+
+    def _consolidate_temp_output_across_obj(
+        self,
+        inference_state,
+        frame_idx,
+        is_cond,
+        run_mem_encoder,
+        consolidate_at_video_res=False,
+    ):
+        """
+        Consolidate the per-object temporary outputs in `temp_output_dict_per_obj` on
+        a frame into a single output for all objects, including
+        1) fill any missing objects either from `output_dict_per_obj` (if they exist in
+           `output_dict_per_obj` for this frame) or leave them as placeholder values
+           (if they don't exist in `output_dict_per_obj` for this frame);
+        2) if specified, rerun memory encoder after apply non-overlapping constraints
+           on the object scores.
+        """
+        batch_size = self._get_obj_num(inference_state)
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+        # Optionally, we allow consolidating the temporary outputs at the original
+        # video resolution (to provide a better editing experience for mask prompts).
+        if consolidate_at_video_res:
+            assert not run_mem_encoder, "memory encoder cannot run at video resolution"
+            consolidated_H = inference_state["video_height"]
+            consolidated_W = inference_state["video_width"]
+            consolidated_mask_key = "pred_masks_video_res"
+        else:
+            consolidated_H = consolidated_W = self.image_size // 4
+            consolidated_mask_key = "pred_masks"
+
+        # Initialize `consolidated_out`. Its "maskmem_features" and "maskmem_pos_enc"
+        # will be added when rerunning the memory encoder after applying non-overlapping
+        # constraints to object scores. Its "pred_masks" are prefilled with a large
+        # negative value (NO_OBJ_SCORE) to represent missing objects.
+        consolidated_out = {
+            "maskmem_features": None,
+            "maskmem_pos_enc": None,
+            consolidated_mask_key: torch.full(
+                size=(batch_size, 1, consolidated_H, consolidated_W),
+                fill_value=NO_OBJ_SCORE,
+                dtype=torch.float32,
+                device=inference_state["storage_device"],
+            ),
+            "obj_ptr": torch.full(
+                size=(batch_size, self.hidden_dim),
+                fill_value=NO_OBJ_SCORE,
+                dtype=torch.float32,
+                device=inference_state["device"],
+            ),
+            "object_score_logits": torch.full(
+                size=(batch_size, 1),
+                # default to 10.0 for object_score_logits, i.e. assuming the object is
+                # present as sigmoid(10)=1, same as in `predict_masks` of `MaskDecoder`
+                fill_value=10.0,
+                dtype=torch.float32,
+                device=inference_state["device"],
+            ),
+        }
+        empty_mask_ptr = None
+        for obj_idx in range(batch_size):
+            obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+            obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+            out = obj_temp_output_dict[storage_key].get(frame_idx, None)
+            # If the object doesn't appear in "temp_output_dict_per_obj" on this frame,
+            # we fall back and look up its previous output in "output_dict_per_obj".
+            # We look up both "cond_frame_outputs" and "non_cond_frame_outputs" in
+            # "output_dict_per_obj" to find a previous output for this object.
+            if out is None:
+                out = obj_output_dict["cond_frame_outputs"].get(frame_idx, None)
+            if out is None:
+                out = obj_output_dict["non_cond_frame_outputs"].get(frame_idx, None)
+            # If the object doesn't appear in "output_dict_per_obj" either, we skip it
+            # and leave its mask scores to the default scores (i.e. the NO_OBJ_SCORE
+            # placeholder above) and set its object pointer to be a dummy pointer.
+            if out is None:
+                # Fill in dummy object pointers for those objects without any inputs or
+                # tracking outcomes on this frame (only do it under `run_mem_encoder=True`,
+                # i.e. when we need to build the memory for tracking).
+                if run_mem_encoder:
+                    if empty_mask_ptr is None:
+                        empty_mask_ptr = self._get_empty_mask_ptr(
+                            inference_state, frame_idx
+                        )
+                    # fill object pointer with a dummy pointer (based on an empty mask)
+                    consolidated_out["obj_ptr"][obj_idx : obj_idx + 1] = empty_mask_ptr
+                continue
+            # Add the temporary object output mask to consolidated output mask
+            obj_mask = out["pred_masks"]
+            consolidated_pred_masks = consolidated_out[consolidated_mask_key]
+            if obj_mask.shape[-2:] == consolidated_pred_masks.shape[-2:]:
+                consolidated_pred_masks[obj_idx : obj_idx + 1] = obj_mask
+            else:
+                # Resize first if temporary object mask has a different resolution
+                resized_obj_mask = torch.nn.functional.interpolate(
+                    obj_mask,
+                    size=consolidated_pred_masks.shape[-2:],
+                    mode="bilinear",
+                    align_corners=False,
+                )
+                consolidated_pred_masks[obj_idx : obj_idx + 1] = resized_obj_mask
+            consolidated_out["obj_ptr"][obj_idx : obj_idx + 1] = out["obj_ptr"]
+            consolidated_out["object_score_logits"][obj_idx : obj_idx + 1] = out[
+                "object_score_logits"
+            ]
+
+        # Optionally, apply non-overlapping constraints on the consolidated scores
+        # and rerun the memory encoder
+        if run_mem_encoder:
+            device = inference_state["device"]
+            high_res_masks = torch.nn.functional.interpolate(
+                consolidated_out["pred_masks"].to(device, non_blocking=True),
+                size=(self.image_size, self.image_size),
+                mode="bilinear",
+                align_corners=False,
+            )
+            if self.non_overlap_masks_for_mem_enc:
+                high_res_masks = self._apply_non_overlapping_constraints(high_res_masks)
+            maskmem_features, maskmem_pos_enc = self._run_memory_encoder(
+                inference_state=inference_state,
+                frame_idx=frame_idx,
+                batch_size=batch_size,
+                high_res_masks=high_res_masks,
+                object_score_logits=consolidated_out["object_score_logits"],
+                is_mask_from_pts=True,  # these frames are what the user interacted with
+            )
+            consolidated_out["maskmem_features"] = maskmem_features
+            consolidated_out["maskmem_pos_enc"] = maskmem_pos_enc
+
+        return consolidated_out
+
+    def _get_empty_mask_ptr(self, inference_state, frame_idx):
+        """Get a dummy object pointer based on an empty mask on the current frame."""
+        # A dummy (empty) mask with a single object
+        batch_size = 1
+        mask_inputs = torch.zeros(
+            (batch_size, 1, self.image_size, self.image_size),
+            dtype=torch.float32,
+            device=inference_state["device"],
+        )
+
+        # Retrieve correct image features
+        (
+            _,
+            _,
+            current_vision_feats,
+            current_vision_pos_embeds,
+            feat_sizes,
+        ) = self._get_image_feature(inference_state, frame_idx, batch_size)
+
+        # Feed the empty mask and image feature above to get a dummy object pointer
+        current_out = self.track_step(
+            frame_idx=frame_idx,
+            is_init_cond_frame=True,
+            current_vision_feats=current_vision_feats,
+            current_vision_pos_embeds=current_vision_pos_embeds,
+            feat_sizes=feat_sizes,
+            point_inputs=None,
+            mask_inputs=mask_inputs,
+            output_dict={},
+            num_frames=inference_state["num_frames"],
+            track_in_reverse=False,
+            run_mem_encoder=False,
+            prev_sam_mask_logits=None,
+        )
+        return current_out["obj_ptr"]
+
+    @torch.inference_mode()
+    def propagate_in_video_preflight(self, inference_state):
+        """Prepare inference_state and consolidate temporary outputs before tracking."""
+        # Tracking has started and we don't allow adding new objects until session is reset.
+        inference_state["tracking_has_started"] = True
+        batch_size = self._get_obj_num(inference_state)
+
+        # Consolidate per-object temporary outputs in "temp_output_dict_per_obj" and
+        # add them into "output_dict".
+        temp_output_dict_per_obj = inference_state["temp_output_dict_per_obj"]
+        output_dict = inference_state["output_dict"]
+        # "consolidated_frame_inds" contains indices of those frames where consolidated
+        # temporary outputs have been added (either in this call or any previous calls
+        # to `propagate_in_video_preflight`).
+        consolidated_frame_inds = inference_state["consolidated_frame_inds"]
+        for is_cond in [False, True]:
+            # Separately consolidate conditioning and non-conditioning temp outputs
+            storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+            # Find all the frames that contain temporary outputs for any objects
+            # (these should be the frames that have just received clicks for mask inputs
+            # via `add_new_points_or_box` or `add_new_mask`)
+            temp_frame_inds = set()
+            for obj_temp_output_dict in temp_output_dict_per_obj.values():
+                temp_frame_inds.update(obj_temp_output_dict[storage_key].keys())
+            consolidated_frame_inds[storage_key].update(temp_frame_inds)
+            # consolidate the temporary output across all objects on this frame
+            for frame_idx in temp_frame_inds:
+                consolidated_out = self._consolidate_temp_output_across_obj(
+                    inference_state, frame_idx, is_cond=is_cond, run_mem_encoder=True
+                )
+                # merge them into "output_dict" and also create per-object slices
+                output_dict[storage_key][frame_idx] = consolidated_out
+                self._add_output_per_object(
+                    inference_state, frame_idx, consolidated_out, storage_key
+                )
+                clear_non_cond_mem = self.clear_non_cond_mem_around_input and (
+                    self.clear_non_cond_mem_for_multi_obj or batch_size <= 1
+                )
+                if clear_non_cond_mem:
+                    # clear non-conditioning memory of the surrounding frames
+                    self._clear_non_cond_mem_around_input(inference_state, frame_idx)
+
+            # clear temporary outputs in `temp_output_dict_per_obj`
+            for obj_temp_output_dict in temp_output_dict_per_obj.values():
+                obj_temp_output_dict[storage_key].clear()
+
+        # edge case: if an output is added to "cond_frame_outputs", we remove any prior
+        # output on the same frame in "non_cond_frame_outputs"
+        for frame_idx in output_dict["cond_frame_outputs"]:
+            output_dict["non_cond_frame_outputs"].pop(frame_idx, None)
+        for obj_output_dict in inference_state["output_dict_per_obj"].values():
+            for frame_idx in obj_output_dict["cond_frame_outputs"]:
+                obj_output_dict["non_cond_frame_outputs"].pop(frame_idx, None)
+        for frame_idx in consolidated_frame_inds["cond_frame_outputs"]:
+            assert frame_idx in output_dict["cond_frame_outputs"]
+            consolidated_frame_inds["non_cond_frame_outputs"].discard(frame_idx)
+
+        # Make sure that the frame indices in "consolidated_frame_inds" are exactly those frames
+        # with either points or mask inputs (which should be true under a correct workflow).
+        all_consolidated_frame_inds = (
+            consolidated_frame_inds["cond_frame_outputs"]
+            | consolidated_frame_inds["non_cond_frame_outputs"]
+        )
+        input_frames_inds = set()
+        for point_inputs_per_frame in inference_state["point_inputs_per_obj"].values():
+            input_frames_inds.update(point_inputs_per_frame.keys())
+        for mask_inputs_per_frame in inference_state["mask_inputs_per_obj"].values():
+            input_frames_inds.update(mask_inputs_per_frame.keys())
+        assert all_consolidated_frame_inds == input_frames_inds
+
+    @torch.inference_mode()
+    def propagate_in_video(
+        self,
+        inference_state,
+        start_frame_idx=None,
+        max_frame_num_to_track=None,
+        reverse=False,
+    ):
+        """Propagate the input points across frames to track in the entire video."""
+        self.propagate_in_video_preflight(inference_state)
+
+        output_dict = inference_state["output_dict"]
+        consolidated_frame_inds = inference_state["consolidated_frame_inds"]
+        obj_ids = inference_state["obj_ids"]
+        num_frames = inference_state["num_frames"]
+        batch_size = self._get_obj_num(inference_state)
+        if len(output_dict["cond_frame_outputs"]) == 0:
+            raise RuntimeError("No points are provided; please add points first")
+        clear_non_cond_mem = self.clear_non_cond_mem_around_input and (
+            self.clear_non_cond_mem_for_multi_obj or batch_size <= 1
+        )
+
+        # set start index, end index, and processing order
+        if start_frame_idx is None:
+            # default: start from the earliest frame with input points
+            start_frame_idx = min(output_dict["cond_frame_outputs"])
+        if max_frame_num_to_track is None:
+            # default: track all the frames in the video
+            max_frame_num_to_track = num_frames
+        if reverse:
+            end_frame_idx = max(start_frame_idx - max_frame_num_to_track, 0)
+            if start_frame_idx > 0:
+                processing_order = range(start_frame_idx, end_frame_idx - 1, -1)
+            else:
+                processing_order = []  # skip reverse tracking if starting from frame 0
+        else:
+            end_frame_idx = min(
+                start_frame_idx + max_frame_num_to_track, num_frames - 1
+            )
+            processing_order = range(start_frame_idx, end_frame_idx + 1)
+
+        for frame_idx in tqdm(processing_order, desc="propagate in video"):
+            # We skip those frames already in consolidated outputs (these are frames
+            # that received input clicks or mask). Note that we cannot directly run
+            # batched forward on them via `_run_single_frame_inference` because the
+            # number of clicks on each object might be different.
+            if frame_idx in consolidated_frame_inds["cond_frame_outputs"]:
+                storage_key = "cond_frame_outputs"
+                current_out = output_dict[storage_key][frame_idx]
+                pred_masks = current_out["pred_masks"]
+                if clear_non_cond_mem:
+                    # clear non-conditioning memory of the surrounding frames
+                    self._clear_non_cond_mem_around_input(inference_state, frame_idx)
+            elif frame_idx in consolidated_frame_inds["non_cond_frame_outputs"]:
+                storage_key = "non_cond_frame_outputs"
+                current_out = output_dict[storage_key][frame_idx]
+                pred_masks = current_out["pred_masks"]
+            else:
+                storage_key = "non_cond_frame_outputs"
+                current_out, pred_masks = self._run_single_frame_inference(
+                    inference_state=inference_state,
+                    output_dict=output_dict,
+                    frame_idx=frame_idx,
+                    batch_size=batch_size,
+                    is_init_cond_frame=False,
+                    point_inputs=None,
+                    mask_inputs=None,
+                    reverse=reverse,
+                    run_mem_encoder=True,
+                )
+                output_dict[storage_key][frame_idx] = current_out
+            # Create slices of per-object outputs for subsequent interaction with each
+            # individual object after tracking.
+            self._add_output_per_object(
+                inference_state, frame_idx, current_out, storage_key
+            )
+            inference_state["frames_already_tracked"][frame_idx] = {"reverse": reverse}
+
+            # Resize the output mask to the original video resolution (we directly use
+            # the mask scores on GPU for output to avoid any CPU conversion in between)
+            _, video_res_masks = self._get_orig_video_res_output(
+                inference_state, pred_masks
+            )
+            yield frame_idx, obj_ids, video_res_masks
+
+    def _add_output_per_object(
+        self, inference_state, frame_idx, current_out, storage_key
+    ):
+        """
+        Split a multi-object output into per-object output slices and add them into
+        `output_dict_per_obj`. The resulting slices share the same tensor storage.
+        """
+        maskmem_features = current_out["maskmem_features"]
+        assert maskmem_features is None or isinstance(maskmem_features, torch.Tensor)
+
+        maskmem_pos_enc = current_out["maskmem_pos_enc"]
+        assert maskmem_pos_enc is None or isinstance(maskmem_pos_enc, list)
+
+        output_dict_per_obj = inference_state["output_dict_per_obj"]
+        for obj_idx, obj_output_dict in output_dict_per_obj.items():
+            obj_slice = slice(obj_idx, obj_idx + 1)
+            obj_out = {
+                "maskmem_features": None,
+                "maskmem_pos_enc": None,
+                "pred_masks": current_out["pred_masks"][obj_slice],
+                "obj_ptr": current_out["obj_ptr"][obj_slice],
+                "object_score_logits": current_out["object_score_logits"][obj_slice],
+            }
+            if maskmem_features is not None:
+                obj_out["maskmem_features"] = maskmem_features[obj_slice]
+            if maskmem_pos_enc is not None:
+                obj_out["maskmem_pos_enc"] = [x[obj_slice] for x in maskmem_pos_enc]
+            obj_output_dict[storage_key][frame_idx] = obj_out
+
+    @torch.inference_mode()
+    def clear_all_prompts_in_frame(
+        self, inference_state, frame_idx, obj_id, need_output=True
+    ):
+        """Remove all input points or mask in a specific frame for a given object."""
+        obj_idx = self._obj_id_to_idx(inference_state, obj_id)
+
+        # Clear the conditioning information on the given frame
+        inference_state["point_inputs_per_obj"][obj_idx].pop(frame_idx, None)
+        inference_state["mask_inputs_per_obj"][obj_idx].pop(frame_idx, None)
+
+        temp_output_dict_per_obj = inference_state["temp_output_dict_per_obj"]
+        temp_output_dict_per_obj[obj_idx]["cond_frame_outputs"].pop(frame_idx, None)
+        temp_output_dict_per_obj[obj_idx]["non_cond_frame_outputs"].pop(frame_idx, None)
+
+        # Check and see if there are still any inputs left on this frame
+        batch_size = self._get_obj_num(inference_state)
+        frame_has_input = False
+        for obj_idx2 in range(batch_size):
+            if frame_idx in inference_state["point_inputs_per_obj"][obj_idx2]:
+                frame_has_input = True
+                break
+            if frame_idx in inference_state["mask_inputs_per_obj"][obj_idx2]:
+                frame_has_input = True
+                break
+
+        # If this frame has no remaining inputs for any objects, we further clear its
+        # conditioning frame status
+        if not frame_has_input:
+            output_dict = inference_state["output_dict"]
+            consolidated_frame_inds = inference_state["consolidated_frame_inds"]
+            consolidated_frame_inds["cond_frame_outputs"].discard(frame_idx)
+            consolidated_frame_inds["non_cond_frame_outputs"].discard(frame_idx)
+            # Remove the frame's conditioning output (possibly downgrading it to non-conditioning)
+            out = output_dict["cond_frame_outputs"].pop(frame_idx, None)
+            if out is not None:
+                # The frame is not a conditioning frame anymore since it's not receiving inputs,
+                # so we "downgrade" its output (if exists) to a non-conditioning frame output.
+                output_dict["non_cond_frame_outputs"][frame_idx] = out
+                inference_state["frames_already_tracked"].pop(frame_idx, None)
+            # Similarly, do it for the sliced output on each object.
+            for obj_idx2 in range(batch_size):
+                obj_output_dict = inference_state["output_dict_per_obj"][obj_idx2]
+                obj_out = obj_output_dict["cond_frame_outputs"].pop(frame_idx, None)
+                if obj_out is not None:
+                    obj_output_dict["non_cond_frame_outputs"][frame_idx] = obj_out
+
+            # If all the conditioning frames have been removed, we also clear the tracking outputs
+            if len(output_dict["cond_frame_outputs"]) == 0:
+                self._reset_tracking_results(inference_state)
+
+        if not need_output:
+            return
+        # Finally, output updated masks per object (after removing the inputs above)
+        obj_ids = inference_state["obj_ids"]
+        is_cond = any(
+            frame_idx in obj_temp_output_dict["cond_frame_outputs"]
+            for obj_temp_output_dict in temp_output_dict_per_obj.values()
+        )
+        consolidated_out = self._consolidate_temp_output_across_obj(
+            inference_state,
+            frame_idx,
+            is_cond=is_cond,
+            run_mem_encoder=False,
+            consolidate_at_video_res=True,
+        )
+        _, video_res_masks = self._get_orig_video_res_output(
+            inference_state, consolidated_out["pred_masks_video_res"]
+        )
+        return frame_idx, obj_ids, video_res_masks
+
+    @torch.inference_mode()
+    def reset_state(self, inference_state):
+        """Remove all input points or mask in all frames throughout the video."""
+        self._reset_tracking_results(inference_state)
+        # Remove all object ids
+        inference_state["obj_id_to_idx"].clear()
+        inference_state["obj_idx_to_id"].clear()
+        inference_state["obj_ids"].clear()
+        inference_state["point_inputs_per_obj"].clear()
+        inference_state["mask_inputs_per_obj"].clear()
+        inference_state["output_dict_per_obj"].clear()
+        inference_state["temp_output_dict_per_obj"].clear()
+
+    def _reset_tracking_results(self, inference_state):
+        """Reset all tracking inputs and results across the videos."""
+        for v in inference_state["point_inputs_per_obj"].values():
+            v.clear()
+        for v in inference_state["mask_inputs_per_obj"].values():
+            v.clear()
+        for v in inference_state["output_dict_per_obj"].values():
+            v["cond_frame_outputs"].clear()
+            v["non_cond_frame_outputs"].clear()
+        for v in inference_state["temp_output_dict_per_obj"].values():
+            v["cond_frame_outputs"].clear()
+            v["non_cond_frame_outputs"].clear()
+        inference_state["output_dict"]["cond_frame_outputs"].clear()
+        inference_state["output_dict"]["non_cond_frame_outputs"].clear()
+        inference_state["consolidated_frame_inds"]["cond_frame_outputs"].clear()
+        inference_state["consolidated_frame_inds"]["non_cond_frame_outputs"].clear()
+        inference_state["tracking_has_started"] = False
+        inference_state["frames_already_tracked"].clear()
+
+    def _get_image_feature(self, inference_state, frame_idx, batch_size):
+        """Compute the image features on a given frame."""
+        # Look up in the cache first
+        image, backbone_out = inference_state["cached_features"].get(
+            frame_idx, (None, None)
+        )
+        if backbone_out is None:
+            # Cache miss -- we will run inference on a single image
+            image = inference_state["images"][frame_idx].to(inference_state["device"]).float().unsqueeze(0)
+            backbone_out = self.forward_image(image)
+            # Cache the most recent frame's feature (for repeated interactions with
+            # a frame; we can use an LRU cache for more frames in the future).
+            inference_state["cached_features"] = {frame_idx: (image, backbone_out)}
+
+        # expand the features to have the same dimension as the number of objects
+        expanded_image = image.expand(batch_size, -1, -1, -1)
+        expanded_backbone_out = {
+            "backbone_fpn": backbone_out["backbone_fpn"].copy(),
+            "vision_pos_enc": backbone_out["vision_pos_enc"].copy(),
+        }
+        for i, feat in enumerate(expanded_backbone_out["backbone_fpn"]):
+            expanded_backbone_out["backbone_fpn"][i] = feat.expand(
+                batch_size, -1, -1, -1
+            )
+        for i, pos in enumerate(expanded_backbone_out["vision_pos_enc"]):
+            pos = pos.expand(batch_size, -1, -1, -1)
+            expanded_backbone_out["vision_pos_enc"][i] = pos
+
+        features = self._prepare_backbone_features(expanded_backbone_out)
+        features = (expanded_image,) + features
+        return features
+
+    def _run_single_frame_inference(
+        self,
+        inference_state,
+        output_dict,
+        frame_idx,
+        batch_size,
+        is_init_cond_frame,
+        point_inputs,
+        mask_inputs,
+        reverse,
+        run_mem_encoder,
+        prev_sam_mask_logits=None,
+    ):
+        """Run tracking on a single frame based on current inputs and previous memory."""
+        # Retrieve correct image features
+        (
+            _,
+            _,
+            current_vision_feats,
+            current_vision_pos_embeds,
+            feat_sizes,
+        ) = self._get_image_feature(inference_state, frame_idx, batch_size)
+
+        # point and mask should not appear as input simultaneously on the same frame
+        assert point_inputs is None or mask_inputs is None
+        current_out = self.track_step(
+            frame_idx=frame_idx,
+            is_init_cond_frame=is_init_cond_frame,
+            current_vision_feats=current_vision_feats,
+            current_vision_pos_embeds=current_vision_pos_embeds,
+            feat_sizes=feat_sizes,
+            point_inputs=point_inputs,
+            mask_inputs=mask_inputs,
+            output_dict=output_dict,
+            num_frames=inference_state["num_frames"],
+            track_in_reverse=reverse,
+            run_mem_encoder=run_mem_encoder,
+            prev_sam_mask_logits=prev_sam_mask_logits,
+        )
+
+        # optionally offload the output to CPU memory to save GPU space
+        storage_device = inference_state["storage_device"]
+        maskmem_features = current_out["maskmem_features"]
+        if maskmem_features is not None:
+            maskmem_features = maskmem_features.to(torch.bfloat16)
+            maskmem_features = maskmem_features.to(storage_device, non_blocking=True)
+        pred_masks_gpu = current_out["pred_masks"]
+        # potentially fill holes in the predicted masks
+        if self.fill_hole_area > 0:
+            pred_masks_gpu = fill_holes_in_mask_scores(
+                pred_masks_gpu, self.fill_hole_area
+            )
+        pred_masks = pred_masks_gpu.to(storage_device, non_blocking=True)
+        # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
+        maskmem_pos_enc = self._get_maskmem_pos_enc(inference_state, current_out)
+        # object pointer is a small tensor, so we always keep it on GPU memory for fast access
+        obj_ptr = current_out["obj_ptr"]
+        object_score_logits = current_out["object_score_logits"]
+        # make a compact version of this frame's output to reduce the state size
+        compact_current_out = {
+            "maskmem_features": maskmem_features,
+            "maskmem_pos_enc": maskmem_pos_enc,
+            "pred_masks": pred_masks,
+            "obj_ptr": obj_ptr,
+            "object_score_logits": object_score_logits,
+        }
+        return compact_current_out, pred_masks_gpu
+
+    def _run_memory_encoder(
+        self,
+        inference_state,
+        frame_idx,
+        batch_size,
+        high_res_masks,
+        object_score_logits,
+        is_mask_from_pts,
+    ):
+        """
+        Run the memory encoder on `high_res_masks`. This is usually after applying
+        non-overlapping constraints to object scores. Since their scores changed, their
+        memory also need to be computed again with the memory encoder.
+        """
+        # Retrieve correct image features
+        _, _, current_vision_feats, _, feat_sizes = self._get_image_feature(
+            inference_state, frame_idx, batch_size
+        )
+        maskmem_features, maskmem_pos_enc = self._encode_new_memory(
+            current_vision_feats=current_vision_feats,
+            feat_sizes=feat_sizes,
+            pred_masks_high_res=high_res_masks,
+            object_score_logits=object_score_logits,
+            is_mask_from_pts=is_mask_from_pts,
+        )
+
+        # optionally offload the output to CPU memory to save GPU space
+        storage_device = inference_state["storage_device"]
+        maskmem_features = maskmem_features.to(torch.bfloat16)
+        maskmem_features = maskmem_features.to(storage_device, non_blocking=True)
+        # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
+        maskmem_pos_enc = self._get_maskmem_pos_enc(
+            inference_state, {"maskmem_pos_enc": maskmem_pos_enc}
+        )
+        return maskmem_features, maskmem_pos_enc
+
+    def _get_maskmem_pos_enc(self, inference_state, current_out):
+        """
+        `maskmem_pos_enc` is the same across frames and objects, so we cache it as
+        a constant in the inference session to reduce session storage size.
+        """
+        model_constants = inference_state["constants"]
+        # "out_maskmem_pos_enc" should be either a list of tensors or None
+        out_maskmem_pos_enc = current_out["maskmem_pos_enc"]
+        if out_maskmem_pos_enc is not None:
+            if "maskmem_pos_enc" not in model_constants:
+                assert isinstance(out_maskmem_pos_enc, list)
+                # only take the slice for one object, since it's same across objects
+                maskmem_pos_enc = [x[0:1].clone() for x in out_maskmem_pos_enc]
+                model_constants["maskmem_pos_enc"] = maskmem_pos_enc
+            else:
+                maskmem_pos_enc = model_constants["maskmem_pos_enc"]
+            # expand the cached maskmem_pos_enc to the actual batch size
+            batch_size = out_maskmem_pos_enc[0].size(0)
+            expanded_maskmem_pos_enc = [
+                x.expand(batch_size, -1, -1, -1) for x in maskmem_pos_enc
+            ]
+        else:
+            expanded_maskmem_pos_enc = None
+        return expanded_maskmem_pos_enc
+
+    @torch.inference_mode()
+    def remove_object(self, inference_state, obj_id, strict=False, need_output=True):
+        """
+        Remove an object id from the tracking state. If strict is True, we check whether
+        the object id actually exists and raise an error if it doesn't exist.
+        """
+        old_obj_idx_to_rm = inference_state["obj_id_to_idx"].get(obj_id, None)
+        updated_frames = []
+        # Check whether this object_id to remove actually exists and possibly raise an error.
+        if old_obj_idx_to_rm is None:
+            if not strict:
+                return inference_state["obj_ids"], updated_frames
+            raise RuntimeError(
+                f"Cannot remove object id {obj_id} as it doesn't exist. "
+                f"All existing object ids: {inference_state['obj_ids']}."
+            )
+
+        # If this is the only remaining object id, we simply reset the state.
+        if len(inference_state["obj_id_to_idx"]) == 1:
+            self.reset_state(inference_state)
+            return inference_state["obj_ids"], updated_frames
+
+        # There are still remaining objects after removing this object id. In this case,
+        # we need to delete the object storage from inference state tensors.
+        # Step 0: clear the input on those frames where this object id has point or mask input
+        # (note that this step is required as it might downgrade conditioning frames to
+        # non-conditioning ones)
+        obj_input_frames_inds = set()
+        obj_input_frames_inds.update(
+            inference_state["point_inputs_per_obj"][old_obj_idx_to_rm]
+        )
+        obj_input_frames_inds.update(
+            inference_state["mask_inputs_per_obj"][old_obj_idx_to_rm]
+        )
+        for frame_idx in obj_input_frames_inds:
+            self.clear_all_prompts_in_frame(
+                inference_state, frame_idx, obj_id, need_output=False
+            )
+
+        # Step 1: Update the object id mapping (note that it must be done after Step 0,
+        # since Step 0 still requires the old object id mappings in inference_state)
+        old_obj_ids = inference_state["obj_ids"]
+        old_obj_inds = list(range(len(old_obj_ids)))
+        remain_old_obj_inds = old_obj_inds.copy()
+        remain_old_obj_inds.remove(old_obj_idx_to_rm)
+        new_obj_ids = [old_obj_ids[old_idx] for old_idx in remain_old_obj_inds]
+        new_obj_inds = list(range(len(new_obj_ids)))
+        # build new mappings
+        old_idx_to_new_idx = dict(zip(remain_old_obj_inds, new_obj_inds))
+        inference_state["obj_id_to_idx"] = dict(zip(new_obj_ids, new_obj_inds))
+        inference_state["obj_idx_to_id"] = dict(zip(new_obj_inds, new_obj_ids))
+        inference_state["obj_ids"] = new_obj_ids
+
+        # Step 2: For per-object tensor storage, we shift their obj_idx in the dict keys.
+        # (note that "consolidated_frame_inds" doesn't need to be updated in this step as
+        # it's already handled in Step 0)
+        def _map_keys(container):
+            new_kvs = []
+            for k in old_obj_inds:
+                v = container.pop(k)
+                if k in old_idx_to_new_idx:
+                    new_kvs.append((old_idx_to_new_idx[k], v))
+            container.update(new_kvs)
+
+        _map_keys(inference_state["point_inputs_per_obj"])
+        _map_keys(inference_state["mask_inputs_per_obj"])
+        _map_keys(inference_state["output_dict_per_obj"])
+        _map_keys(inference_state["temp_output_dict_per_obj"])
+
+        # Step 3: For packed tensor storage, we index the remaining ids and rebuild the per-object slices.
+        def _slice_state(output_dict, storage_key):
+            for frame_idx, out in output_dict[storage_key].items():
+                out["maskmem_features"] = out["maskmem_features"][remain_old_obj_inds]
+                out["maskmem_pos_enc"] = [
+                    x[remain_old_obj_inds] for x in out["maskmem_pos_enc"]
+                ]
+                # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
+                out["maskmem_pos_enc"] = self._get_maskmem_pos_enc(inference_state, out)
+                out["pred_masks"] = out["pred_masks"][remain_old_obj_inds]
+                out["obj_ptr"] = out["obj_ptr"][remain_old_obj_inds]
+                out["object_score_logits"] = out["object_score_logits"][
+                    remain_old_obj_inds
+                ]
+                # also update the per-object slices
+                self._add_output_per_object(
+                    inference_state, frame_idx, out, storage_key
+                )
+
+        _slice_state(inference_state["output_dict"], "cond_frame_outputs")
+        _slice_state(inference_state["output_dict"], "non_cond_frame_outputs")
+
+        # Step 4: Further collect the outputs on those frames in `obj_input_frames_inds`, which
+        # could show an updated mask for objects previously occluded by the object being removed
+        if need_output:
+            temp_output_dict_per_obj = inference_state["temp_output_dict_per_obj"]
+            for frame_idx in obj_input_frames_inds:
+                is_cond = any(
+                    frame_idx in obj_temp_output_dict["cond_frame_outputs"]
+                    for obj_temp_output_dict in temp_output_dict_per_obj.values()
+                )
+                consolidated_out = self._consolidate_temp_output_across_obj(
+                    inference_state,
+                    frame_idx,
+                    is_cond=is_cond,
+                    run_mem_encoder=False,
+                    consolidate_at_video_res=True,
+                )
+                _, video_res_masks = self._get_orig_video_res_output(
+                    inference_state, consolidated_out["pred_masks_video_res"]
+                )
+                updated_frames.append((frame_idx, video_res_masks))
+
+        return inference_state["obj_ids"], updated_frames
+
+    def _clear_non_cond_mem_around_input(self, inference_state, frame_idx):
+        """
+        Remove the non-conditioning memory around the input frame. When users provide
+        correction clicks, the surrounding frames' non-conditioning memories can still
+        contain outdated object appearance information and could confuse the model.
+
+        This method clears those non-conditioning memories surrounding the interacted
+        frame to avoid giving the model both old and new information about the object.
+        """
+        r = self.memory_temporal_stride_for_eval
+        frame_idx_begin = frame_idx - r * self.num_maskmem
+        frame_idx_end = frame_idx + r * self.num_maskmem
+        output_dict = inference_state["output_dict"]
+        non_cond_frame_outputs = output_dict["non_cond_frame_outputs"]
+        for t in range(frame_idx_begin, frame_idx_end + 1):
+            non_cond_frame_outputs.pop(t, None)
+            for obj_output_dict in inference_state["output_dict_per_obj"].values():
+                obj_output_dict["non_cond_frame_outputs"].pop(t, None)

custom_nodes/comfyui-segment-anything-2/sam2/utils/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

custom_nodes/comfyui-segment-anything-2/sam2/utils/amg.py

@@ -0,0 +1,348 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from copy import deepcopy
+from itertools import product
+from typing import Any, Dict, Generator, ItemsView, List, Tuple
+
+import numpy as np
+import torch
+
+# Very lightly adapted from https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/utils/amg.py
+
+
+class MaskData:
+    """
+    A structure for storing masks and their related data in batched format.
+    Implements basic filtering and concatenation.
+    """
+
+    def __init__(self, **kwargs) -> None:
+        for v in kwargs.values():
+            assert isinstance(
+                v, (list, np.ndarray, torch.Tensor)
+            ), "MaskData only supports list, numpy arrays, and torch tensors."
+        self._stats = dict(**kwargs)
+
+    def __setitem__(self, key: str, item: Any) -> None:
+        assert isinstance(
+            item, (list, np.ndarray, torch.Tensor)
+        ), "MaskData only supports list, numpy arrays, and torch tensors."
+        self._stats[key] = item
+
+    def __delitem__(self, key: str) -> None:
+        del self._stats[key]
+
+    def __getitem__(self, key: str) -> Any:
+        return self._stats[key]
+
+    def items(self) -> ItemsView[str, Any]:
+        return self._stats.items()
+
+    def filter(self, keep: torch.Tensor) -> None:
+        for k, v in self._stats.items():
+            if v is None:
+                self._stats[k] = None
+            elif isinstance(v, torch.Tensor):
+                self._stats[k] = v[torch.as_tensor(keep, device=v.device)]
+            elif isinstance(v, np.ndarray):
+                self._stats[k] = v[keep.detach().cpu().numpy()]
+            elif isinstance(v, list) and keep.dtype == torch.bool:
+                self._stats[k] = [a for i, a in enumerate(v) if keep[i]]
+            elif isinstance(v, list):
+                self._stats[k] = [v[i] for i in keep]
+            else:
+                raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
+
+    def cat(self, new_stats: "MaskData") -> None:
+        for k, v in new_stats.items():
+            if k not in self._stats or self._stats[k] is None:
+                self._stats[k] = deepcopy(v)
+            elif isinstance(v, torch.Tensor):
+                self._stats[k] = torch.cat([self._stats[k], v], dim=0)
+            elif isinstance(v, np.ndarray):
+                self._stats[k] = np.concatenate([self._stats[k], v], axis=0)
+            elif isinstance(v, list):
+                self._stats[k] = self._stats[k] + deepcopy(v)
+            else:
+                raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
+
+    def to_numpy(self) -> None:
+        for k, v in self._stats.items():
+            if isinstance(v, torch.Tensor):
+                self._stats[k] = v.float().detach().cpu().numpy()
+
+
+def is_box_near_crop_edge(
+    boxes: torch.Tensor, crop_box: List[int], orig_box: List[int], atol: float = 20.0
+) -> torch.Tensor:
+    """Filter masks at the edge of a crop, but not at the edge of the original image."""
+    crop_box_torch = torch.as_tensor(crop_box, dtype=torch.float, device=boxes.device)
+    orig_box_torch = torch.as_tensor(orig_box, dtype=torch.float, device=boxes.device)
+    boxes = uncrop_boxes_xyxy(boxes, crop_box).float()
+    near_crop_edge = torch.isclose(boxes, crop_box_torch[None, :], atol=atol, rtol=0)
+    near_image_edge = torch.isclose(boxes, orig_box_torch[None, :], atol=atol, rtol=0)
+    near_crop_edge = torch.logical_and(near_crop_edge, ~near_image_edge)
+    return torch.any(near_crop_edge, dim=1)
+
+
+def box_xyxy_to_xywh(box_xyxy: torch.Tensor) -> torch.Tensor:
+    box_xywh = deepcopy(box_xyxy)
+    box_xywh[2] = box_xywh[2] - box_xywh[0]
+    box_xywh[3] = box_xywh[3] - box_xywh[1]
+    return box_xywh
+
+
+def batch_iterator(batch_size: int, *args) -> Generator[List[Any], None, None]:
+    assert len(args) > 0 and all(
+        len(a) == len(args[0]) for a in args
+    ), "Batched iteration must have inputs of all the same size."
+    n_batches = len(args[0]) // batch_size + int(len(args[0]) % batch_size != 0)
+    for b in range(n_batches):
+        yield [arg[b * batch_size : (b + 1) * batch_size] for arg in args]
+
+
+def mask_to_rle_pytorch(tensor: torch.Tensor) -> List[Dict[str, Any]]:
+    """
+    Encodes masks to an uncompressed RLE, in the format expected by
+    pycoco tools.
+    """
+    # Put in fortran order and flatten h,w
+    b, h, w = tensor.shape
+    tensor = tensor.permute(0, 2, 1).flatten(1)
+
+    # Compute change indices
+    diff = tensor[:, 1:] ^ tensor[:, :-1]
+    change_indices = diff.nonzero()
+
+    # Encode run length
+    out = []
+    for i in range(b):
+        cur_idxs = change_indices[change_indices[:, 0] == i, 1]
+        cur_idxs = torch.cat(
+            [
+                torch.tensor([0], dtype=cur_idxs.dtype, device=cur_idxs.device),
+                cur_idxs + 1,
+                torch.tensor([h * w], dtype=cur_idxs.dtype, device=cur_idxs.device),
+            ]
+        )
+        btw_idxs = cur_idxs[1:] - cur_idxs[:-1]
+        counts = [] if tensor[i, 0] == 0 else [0]
+        counts.extend(btw_idxs.detach().cpu().tolist())
+        out.append({"size": [h, w], "counts": counts})
+    return out
+
+
+def rle_to_mask(rle: Dict[str, Any]) -> np.ndarray:
+    """Compute a binary mask from an uncompressed RLE."""
+    h, w = rle["size"]
+    mask = np.empty(h * w, dtype=bool)
+    idx = 0
+    parity = False
+    for count in rle["counts"]:
+        mask[idx : idx + count] = parity
+        idx += count
+        parity ^= True
+    mask = mask.reshape(w, h)
+    return mask.transpose()  # Put in C order
+
+
+def area_from_rle(rle: Dict[str, Any]) -> int:
+    return sum(rle["counts"][1::2])
+
+
+def calculate_stability_score(
+    masks: torch.Tensor, mask_threshold: float, threshold_offset: float
+) -> torch.Tensor:
+    """
+    Computes the stability score for a batch of masks. The stability
+    score is the IoU between the binary masks obtained by thresholding
+    the predicted mask logits at high and low values.
+    """
+    # One mask is always contained inside the other.
+    # Save memory by preventing unnecessary cast to torch.int64
+    intersections = (
+        (masks > (mask_threshold + threshold_offset))
+        .sum(-1, dtype=torch.int16)
+        .sum(-1, dtype=torch.int32)
+    )
+    unions = (
+        (masks > (mask_threshold - threshold_offset))
+        .sum(-1, dtype=torch.int16)
+        .sum(-1, dtype=torch.int32)
+    )
+    return intersections / unions
+
+
+def build_point_grid(n_per_side: int) -> np.ndarray:
+    """Generates a 2D grid of points evenly spaced in [0,1]x[0,1]."""
+    offset = 1 / (2 * n_per_side)
+    points_one_side = np.linspace(offset, 1 - offset, n_per_side)
+    points_x = np.tile(points_one_side[None, :], (n_per_side, 1))
+    points_y = np.tile(points_one_side[:, None], (1, n_per_side))
+    points = np.stack([points_x, points_y], axis=-1).reshape(-1, 2)
+    return points
+
+
+def build_all_layer_point_grids(
+    n_per_side: int, n_layers: int, scale_per_layer: int
+) -> List[np.ndarray]:
+    """Generates point grids for all crop layers."""
+    points_by_layer = []
+    for i in range(n_layers + 1):
+        n_points = int(n_per_side / (scale_per_layer**i))
+        points_by_layer.append(build_point_grid(n_points))
+    return points_by_layer
+
+
+def generate_crop_boxes(
+    im_size: Tuple[int, ...], n_layers: int, overlap_ratio: float
+) -> Tuple[List[List[int]], List[int]]:
+    """
+    Generates a list of crop boxes of different sizes. Each layer
+    has (2**i)**2 boxes for the ith layer.
+    """
+    crop_boxes, layer_idxs = [], []
+    im_h, im_w = im_size
+    short_side = min(im_h, im_w)
+
+    # Original image
+    crop_boxes.append([0, 0, im_w, im_h])
+    layer_idxs.append(0)
+
+    def crop_len(orig_len, n_crops, overlap):
+        return int(math.ceil((overlap * (n_crops - 1) + orig_len) / n_crops))
+
+    for i_layer in range(n_layers):
+        n_crops_per_side = 2 ** (i_layer + 1)
+        overlap = int(overlap_ratio * short_side * (2 / n_crops_per_side))
+
+        crop_w = crop_len(im_w, n_crops_per_side, overlap)
+        crop_h = crop_len(im_h, n_crops_per_side, overlap)
+
+        crop_box_x0 = [int((crop_w - overlap) * i) for i in range(n_crops_per_side)]
+        crop_box_y0 = [int((crop_h - overlap) * i) for i in range(n_crops_per_side)]
+
+        # Crops in XYWH format
+        for x0, y0 in product(crop_box_x0, crop_box_y0):
+            box = [x0, y0, min(x0 + crop_w, im_w), min(y0 + crop_h, im_h)]
+            crop_boxes.append(box)
+            layer_idxs.append(i_layer + 1)
+
+    return crop_boxes, layer_idxs
+
+
+def uncrop_boxes_xyxy(boxes: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
+    x0, y0, _, _ = crop_box
+    offset = torch.tensor([[x0, y0, x0, y0]], device=boxes.device)
+    # Check if boxes has a channel dimension
+    if len(boxes.shape) == 3:
+        offset = offset.unsqueeze(1)
+    return boxes + offset
+
+
+def uncrop_points(points: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
+    x0, y0, _, _ = crop_box
+    offset = torch.tensor([[x0, y0]], device=points.device)
+    # Check if points has a channel dimension
+    if len(points.shape) == 3:
+        offset = offset.unsqueeze(1)
+    return points + offset
+
+
+def uncrop_masks(
+    masks: torch.Tensor, crop_box: List[int], orig_h: int, orig_w: int
+) -> torch.Tensor:
+    x0, y0, x1, y1 = crop_box
+    if x0 == 0 and y0 == 0 and x1 == orig_w and y1 == orig_h:
+        return masks
+    # Coordinate transform masks
+    pad_x, pad_y = orig_w - (x1 - x0), orig_h - (y1 - y0)
+    pad = (x0, pad_x - x0, y0, pad_y - y0)
+    return torch.nn.functional.pad(masks, pad, value=0)
+
+
+def remove_small_regions(
+    mask: np.ndarray, area_thresh: float, mode: str
+) -> Tuple[np.ndarray, bool]:
+    """
+    Removes small disconnected regions and holes in a mask. Returns the
+    mask and an indicator of if the mask has been modified.
+    """
+    import cv2  # type: ignore
+
+    assert mode in ["holes", "islands"]
+    correct_holes = mode == "holes"
+    working_mask = (correct_holes ^ mask).astype(np.uint8)
+    n_labels, regions, stats, _ = cv2.connectedComponentsWithStats(working_mask, 8)
+    sizes = stats[:, -1][1:]  # Row 0 is background label
+    small_regions = [i + 1 for i, s in enumerate(sizes) if s < area_thresh]
+    if len(small_regions) == 0:
+        return mask, False
+    fill_labels = [0] + small_regions
+    if not correct_holes:
+        fill_labels = [i for i in range(n_labels) if i not in fill_labels]
+        # If every region is below threshold, keep largest
+        if len(fill_labels) == 0:
+            fill_labels = [int(np.argmax(sizes)) + 1]
+    mask = np.isin(regions, fill_labels)
+    return mask, True
+
+
+def coco_encode_rle(uncompressed_rle: Dict[str, Any]) -> Dict[str, Any]:
+    from pycocotools import mask as mask_utils  # type: ignore
+
+    h, w = uncompressed_rle["size"]
+    rle = mask_utils.frPyObjects(uncompressed_rle, h, w)
+    rle["counts"] = rle["counts"].decode("utf-8")  # Necessary to serialize with json
+    return rle
+
+
+def batched_mask_to_box(masks: torch.Tensor) -> torch.Tensor:
+    """
+    Calculates boxes in XYXY format around masks. Return [0,0,0,0] for
+    an empty mask. For input shape C1xC2x...xHxW, the output shape is C1xC2x...x4.
+    """
+    # torch.max below raises an error on empty inputs, just skip in this case
+    if torch.numel(masks) == 0:
+        return torch.zeros(*masks.shape[:-2], 4, device=masks.device)
+
+    # Normalize shape to CxHxW
+    shape = masks.shape
+    h, w = shape[-2:]
+    if len(shape) > 2:
+        masks = masks.flatten(0, -3)
+    else:
+        masks = masks.unsqueeze(0)
+
+    # Get top and bottom edges
+    in_height, _ = torch.max(masks, dim=-1)
+    in_height_coords = in_height * torch.arange(h, device=in_height.device)[None, :]
+    bottom_edges, _ = torch.max(in_height_coords, dim=-1)
+    in_height_coords = in_height_coords + h * (~in_height)
+    top_edges, _ = torch.min(in_height_coords, dim=-1)
+
+    # Get left and right edges
+    in_width, _ = torch.max(masks, dim=-2)
+    in_width_coords = in_width * torch.arange(w, device=in_width.device)[None, :]
+    right_edges, _ = torch.max(in_width_coords, dim=-1)
+    in_width_coords = in_width_coords + w * (~in_width)
+    left_edges, _ = torch.min(in_width_coords, dim=-1)
+
+    # If the mask is empty the right edge will be to the left of the left edge.
+    # Replace these boxes with [0, 0, 0, 0]
+    empty_filter = (right_edges < left_edges) | (bottom_edges < top_edges)
+    out = torch.stack([left_edges, top_edges, right_edges, bottom_edges], dim=-1)
+    out = out * (~empty_filter).unsqueeze(-1)
+
+    # Return to original shape
+    if len(shape) > 2:
+        out = out.reshape(*shape[:-2], 4)
+    else:
+        out = out[0]
+
+    return out

custom_nodes/comfyui-segment-anything-2/sam2/utils/misc.py

@@ -0,0 +1,349 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import warnings
+from threading import Thread
+
+import numpy as np
+import torch
+from PIL import Image
+from tqdm import tqdm
+import platform
+
+def get_sdpa_settings():
+    if torch.cuda.is_available():
+        old_gpu = torch.cuda.get_device_properties(0).major < 7
+        # only use Flash Attention on Ampere (8.0) or newer GPUs
+        use_flash_attn = torch.cuda.get_device_properties(0).major >= 8 and platform.system() == 'Linux'
+        # if not use_flash_attn:
+        #     warnings.warn(
+        #         "Flash Attention is disabled as it requires a GPU with Ampere (8.0) CUDA capability.",
+        #         category=UserWarning,
+        #         stacklevel=2,
+        #     )
+        # keep math kernel for PyTorch versions before 2.2 (Flash Attention v2 is only
+        # available on PyTorch 2.2+, while Flash Attention v1 cannot handle all cases)
+        pytorch_version = tuple(int(v) for v in torch.__version__.split(".")[:2])
+        if pytorch_version < (2, 2):
+            warnings.warn(
+                f"You are using PyTorch {torch.__version__} without Flash Attention v2 support. "
+                "Consider upgrading to PyTorch 2.2+ for Flash Attention v2 (which could be faster).",
+                category=UserWarning,
+                stacklevel=2,
+            )
+        math_kernel_on = pytorch_version < (2, 2) or not use_flash_attn
+    else:
+        old_gpu = True
+        use_flash_attn = False
+        math_kernel_on = True
+
+    return old_gpu, use_flash_attn, math_kernel_on
+
+
+def get_connected_components(mask):
+    """
+    Get the connected components (8-connectivity) of binary masks of shape (N, 1, H, W).
+
+    Inputs:
+    - mask: A binary mask tensor of shape (N, 1, H, W), where 1 is foreground and 0 is
+            background.
+
+    Outputs:
+    - labels: A tensor of shape (N, 1, H, W) containing the connected component labels
+              for foreground pixels and 0 for background pixels.
+    - counts: A tensor of shape (N, 1, H, W) containing the area of the connected
+              components for foreground pixels and 0 for background pixels.
+    """
+    from ...sam2 import _C
+
+    return _C.get_connected_componnets(mask.to(torch.uint8).contiguous())
+
+
+def mask_to_box(masks: torch.Tensor):
+    """
+    compute bounding box given an input mask
+
+    Inputs:
+    - masks: [B, 1, H, W] masks, dtype=torch.Tensor
+
+    Returns:
+    - box_coords: [B, 1, 4], contains (x, y) coordinates of top left and bottom right box corners, dtype=torch.Tensor
+    """
+    B, _, h, w = masks.shape
+    device = masks.device
+    xs = torch.arange(w, device=device, dtype=torch.int32)
+    ys = torch.arange(h, device=device, dtype=torch.int32)
+    grid_xs, grid_ys = torch.meshgrid(xs, ys, indexing="xy")
+    grid_xs = grid_xs[None, None, ...].expand(B, 1, h, w)
+    grid_ys = grid_ys[None, None, ...].expand(B, 1, h, w)
+    min_xs, _ = torch.min(torch.where(masks, grid_xs, w).flatten(-2), dim=-1)
+    max_xs, _ = torch.max(torch.where(masks, grid_xs, -1).flatten(-2), dim=-1)
+    min_ys, _ = torch.min(torch.where(masks, grid_ys, h).flatten(-2), dim=-1)
+    max_ys, _ = torch.max(torch.where(masks, grid_ys, -1).flatten(-2), dim=-1)
+    bbox_coords = torch.stack((min_xs, min_ys, max_xs, max_ys), dim=-1)
+
+    return bbox_coords
+
+
+def _load_img_as_tensor(img_path, image_size):
+    img_pil = Image.open(img_path)
+    img_np = np.array(img_pil.convert("RGB").resize((image_size, image_size)))
+    if img_np.dtype == np.uint8:  # np.uint8 is expected for JPEG images
+        img_np = img_np / 255.0
+    else:
+        raise RuntimeError(f"Unknown image dtype: {img_np.dtype} on {img_path}")
+    img = torch.from_numpy(img_np).permute(2, 0, 1)
+    video_width, video_height = img_pil.size  # the original video size
+    return img, video_height, video_width
+
+
+class AsyncVideoFrameLoader:
+    """
+    A list of video frames to be load asynchronously without blocking session start.
+    """
+
+    def __init__(
+        self,
+        img_paths,
+        image_size,
+        offload_video_to_cpu,
+        img_mean,
+        img_std,
+        compute_device,
+    ):
+        self.img_paths = img_paths
+        self.image_size = image_size
+        self.offload_video_to_cpu = offload_video_to_cpu
+        self.img_mean = img_mean
+        self.img_std = img_std
+        # items in `self.images` will be loaded asynchronously
+        self.images = [None] * len(img_paths)
+        # catch and raise any exceptions in the async loading thread
+        self.exception = None
+        # video_height and video_width be filled when loading the first image
+        self.video_height = None
+        self.video_width = None
+        self.compute_device = compute_device
+
+        # load the first frame to fill video_height and video_width and also
+        # to cache it (since it's most likely where the user will click)
+        self.__getitem__(0)
+
+        # load the rest of frames asynchronously without blocking the session start
+        def _load_frames():
+            try:
+                for n in tqdm(range(len(self.images)), desc="frame loading (JPEG)"):
+                    self.__getitem__(n)
+            except Exception as e:
+                self.exception = e
+
+        self.thread = Thread(target=_load_frames, daemon=True)
+        self.thread.start()
+
+    def __getitem__(self, index):
+        if self.exception is not None:
+            raise RuntimeError("Failure in frame loading thread") from self.exception
+
+        img = self.images[index]
+        if img is not None:
+            return img
+
+        img, video_height, video_width = _load_img_as_tensor(
+            self.img_paths[index], self.image_size
+        )
+        self.video_height = video_height
+        self.video_width = video_width
+        # normalize by mean and std
+        img -= self.img_mean
+        img /= self.img_std
+        if not self.offload_video_to_cpu:
+            img = img.to(self.compute_device, non_blocking=True)
+        self.images[index] = img
+        return img
+
+    def __len__(self):
+        return len(self.images)
+
+
+def load_video_frames(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.485, 0.456, 0.406),
+    img_std=(0.229, 0.224, 0.225),
+    async_loading_frames=False,
+    compute_device=torch.device("cuda"),
+):
+    """
+    Load the video frames from video_path. The frames are resized to image_size as in
+    the model and are loaded to GPU if offload_video_to_cpu=False. This is used by the demo.
+    """
+    is_bytes = isinstance(video_path, bytes)
+    is_str = isinstance(video_path, str)
+    is_mp4_path = is_str and os.path.splitext(video_path)[-1] in [".mp4", ".MP4"]
+    if is_bytes or is_mp4_path:
+        return load_video_frames_from_video_file(
+            video_path=video_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            compute_device=compute_device,
+        )
+    elif is_str and os.path.isdir(video_path):
+        return load_video_frames_from_jpg_images(
+            video_path=video_path,
+            image_size=image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            img_mean=img_mean,
+            img_std=img_std,
+            async_loading_frames=async_loading_frames,
+            compute_device=compute_device,
+        )
+    else:
+        raise NotImplementedError(
+            "Only MP4 video and JPEG folder are supported at this moment"
+        )
+
+
+def load_video_frames_from_jpg_images(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.485, 0.456, 0.406),
+    img_std=(0.229, 0.224, 0.225),
+    async_loading_frames=False,
+    compute_device=torch.device("cuda"),
+):
+    """
+    Load the video frames from a directory of JPEG files ("<frame_index>.jpg" format).
+
+    The frames are resized to image_size x image_size and are loaded to GPU if
+    `offload_video_to_cpu` is `False` and to CPU if `offload_video_to_cpu` is `True`.
+
+    You can load a frame asynchronously by setting `async_loading_frames` to `True`.
+    """
+    if isinstance(video_path, str) and os.path.isdir(video_path):
+        jpg_folder = video_path
+    else:
+        raise NotImplementedError(
+            "Only JPEG frames are supported at this moment. For video files, you may use "
+            "ffmpeg (https://ffmpeg.org/) to extract frames into a folder of JPEG files, such as \n"
+            "```\n"
+            "ffmpeg -i <your_video>.mp4 -q:v 2 -start_number 0 <output_dir>/'%05d.jpg'\n"
+            "```\n"
+            "where `-q:v` generates high-quality JPEG frames and `-start_number 0` asks "
+            "ffmpeg to start the JPEG file from 00000.jpg."
+        )
+
+    frame_names = [
+        p
+        for p in os.listdir(jpg_folder)
+        if os.path.splitext(p)[-1] in [".jpg", ".jpeg", ".JPG", ".JPEG"]
+    ]
+    frame_names.sort(key=lambda p: int(os.path.splitext(p)[0]))
+    num_frames = len(frame_names)
+    if num_frames == 0:
+        raise RuntimeError(f"no images found in {jpg_folder}")
+    img_paths = [os.path.join(jpg_folder, frame_name) for frame_name in frame_names]
+    img_mean = torch.tensor(img_mean, dtype=torch.float32)[:, None, None]
+    img_std = torch.tensor(img_std, dtype=torch.float32)[:, None, None]
+
+    if async_loading_frames:
+        lazy_images = AsyncVideoFrameLoader(
+            img_paths,
+            image_size,
+            offload_video_to_cpu,
+            img_mean,
+            img_std,
+            compute_device,
+        )
+        return lazy_images, lazy_images.video_height, lazy_images.video_width
+
+    images = torch.zeros(num_frames, 3, image_size, image_size, dtype=torch.float32)
+    for n, img_path in enumerate(tqdm(img_paths, desc="frame loading (JPEG)")):
+        images[n], video_height, video_width = _load_img_as_tensor(img_path, image_size)
+    if not offload_video_to_cpu:
+        images = images.to(compute_device)
+        img_mean = img_mean.to(compute_device)
+        img_std = img_std.to(compute_device)
+    # normalize by mean and std
+    images -= img_mean
+    images /= img_std
+    return images, video_height, video_width
+
+
+def load_video_frames_from_video_file(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.485, 0.456, 0.406),
+    img_std=(0.229, 0.224, 0.225),
+    compute_device=torch.device("cuda"),
+):
+    """Load the video frames from a video file."""
+    import decord
+
+    img_mean = torch.tensor(img_mean, dtype=torch.float32)[:, None, None]
+    img_std = torch.tensor(img_std, dtype=torch.float32)[:, None, None]
+    # Get the original video height and width
+    decord.bridge.set_bridge("torch")
+    video_height, video_width, _ = decord.VideoReader(video_path).next().shape
+    # Iterate over all frames in the video
+    images = []
+    for frame in decord.VideoReader(video_path, width=image_size, height=image_size):
+        images.append(frame.permute(2, 0, 1))
+
+    images = torch.stack(images, dim=0).float() / 255.0
+    if not offload_video_to_cpu:
+        images = images.to(compute_device)
+        img_mean = img_mean.to(compute_device)
+        img_std = img_std.to(compute_device)
+    # normalize by mean and std
+    images -= img_mean
+    images /= img_std
+    return images, video_height, video_width
+
+
+def fill_holes_in_mask_scores(mask, max_area):
+    """
+    A post processor to fill small holes in mask scores with area under `max_area`.
+    """
+    # Holes are those connected components in background with area <= self.max_area
+    # (background regions are those with mask scores <= 0)
+    assert max_area > 0, "max_area must be positive"
+
+    input_mask = mask
+    try:
+        labels, areas = get_connected_components(mask <= 0)
+        is_hole = (labels > 0) & (areas <= max_area)
+        # We fill holes with a small positive mask score (0.1) to change them to foreground.
+        mask = torch.where(is_hole, 0.1, mask)
+    except Exception as e:
+        # Skip the post-processing step on removing small holes if the CUDA kernel fails
+        warnings.warn(
+            f"{e}\n\nSkipping the post-processing step due to the error above. You can "
+            "still use SAM 2 and it's OK to ignore the error above, although some post-processing "
+            "functionality may be limited (which doesn't affect the results in most cases; see "
+            "https://github.com/facebookresearch/sam2/blob/main/INSTALL.md).",
+            category=UserWarning,
+            stacklevel=2,
+        )
+        mask = input_mask
+
+    return mask
+
+
+def concat_points(old_point_inputs, new_points, new_labels):
+    """Add new points and labels to previous point inputs (add at the end)."""
+    if old_point_inputs is None:
+        points, labels = new_points, new_labels
+    else:
+        points = torch.cat([old_point_inputs["point_coords"], new_points], dim=1)
+        labels = torch.cat([old_point_inputs["point_labels"], new_labels], dim=1)
+
+    return {"point_coords": points, "point_labels": labels}

custom_nodes/comfyui-segment-anything-2/sam2/utils/transforms.py

@@ -0,0 +1,106 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.transforms import Normalize, Resize, ToTensor
+
+
+class SAM2Transforms(nn.Module):
+    def __init__(
+        self, resolution, mask_threshold, max_hole_area=0.0, max_sprinkle_area=0.0
+    ):
+        """
+        Transforms for SAM2.
+        """
+        super().__init__()
+        self.resolution = resolution
+        self.mask_threshold = mask_threshold
+        self.max_hole_area = max_hole_area
+        self.max_sprinkle_area = max_sprinkle_area
+        self.mean = [0.485, 0.456, 0.406]
+        self.std = [0.229, 0.224, 0.225]
+        self.to_tensor = ToTensor()
+        try:
+            self.transforms = torch.jit.script(
+                nn.Sequential(
+                    Resize((self.resolution, self.resolution)),
+                    Normalize(self.mean, self.std),
+                )
+                )
+        except Exception as e:
+            print(f"Failed to torch jit script transforms: {e}, falling back to normal transforms")
+            self.transforms = nn.Sequential(
+                Resize((self.resolution, self.resolution)),
+                Normalize(self.mean, self.std),
+        )
+
+    def __call__(self, x):
+        x = self.to_tensor(x)
+        return self.transforms(x)
+
+    def forward_batch(self, img_list):
+        img_batch = [self.transforms(self.to_tensor(img)) for img in img_list]
+        img_batch = torch.stack(img_batch, dim=0)
+        return img_batch
+
+    def transform_coords(
+        self, coords: torch.Tensor, normalize=False, orig_hw=None
+    ) -> torch.Tensor:
+        """
+        Expects a torch tensor with length 2 in the last dimension. The coordinates can be in absolute image or normalized coordinates,
+        If the coords are in absolute image coordinates, normalize should be set to True and original image size is required.
+
+        Returns
+            Un-normalized coordinates in the range of [0, 1] which is expected by the SAM2 model.
+        """
+        if normalize:
+            assert orig_hw is not None
+            h, w = orig_hw
+            coords = coords.clone()
+            coords[..., 0] = coords[..., 0] / w
+            coords[..., 1] = coords[..., 1] / h
+
+        coords = coords * self.resolution  # unnormalize coords
+        return coords
+
+    def transform_boxes(
+        self, boxes: torch.Tensor, normalize=False, orig_hw=None
+    ) -> torch.Tensor:
+        """
+        Expects a tensor of shape Bx4. The coordinates can be in absolute image or normalized coordinates,
+        if the coords are in absolute image coordinates, normalize should be set to True and original image size is required.
+        """
+        boxes = self.transform_coords(boxes.reshape(-1, 2, 2), normalize, orig_hw)
+        return boxes
+
+    def postprocess_masks(self, masks: torch.Tensor, orig_hw) -> torch.Tensor:
+        """
+        Perform PostProcessing on output masks.
+        """
+        #from ...sam2.utils.misc import get_connected_components
+
+        masks = masks.float()
+        # if self.max_hole_area > 0:
+        #     # Holes are those connected components in background with area <= self.fill_hole_area
+        #     # (background regions are those with mask scores <= self.mask_threshold)
+        #     mask_flat = masks.flatten(0, 1).unsqueeze(1)  # flatten as 1-channel image
+        #     labels, areas = get_connected_components(mask_flat <= self.mask_threshold)
+        #     is_hole = (labels > 0) & (areas <= self.max_hole_area)
+        #     is_hole = is_hole.reshape_as(masks)
+        #     # We fill holes with a small positive mask score (10.0) to change them to foreground.
+        #     masks = torch.where(is_hole, self.mask_threshold + 10.0, masks)
+
+        # if self.max_sprinkle_area > 0:
+        #     labels, areas = get_connected_components(mask_flat > self.mask_threshold)
+        #     is_hole = (labels > 0) & (areas <= self.max_sprinkle_area)
+        #     is_hole = is_hole.reshape_as(masks)
+        #     # We fill holes with negative mask score (-10.0) to change them to background.
+        #     masks = torch.where(is_hole, self.mask_threshold - 10.0, masks)
+
+        masks = F.interpolate(masks, orig_hw, mode="bilinear", align_corners=False)
+        return masks

custom_nodes/comfyui-segment-anything-2/sam2_configs/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

custom_nodes/comfyui-segment-anything-2/sam2_configs/sam2.1_hiera_b+.yaml

@@ -0,0 +1,116 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 112
+      num_heads: 2
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [896, 448, 224, 112]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  no_obj_embed_spatial: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: true
+  proj_tpos_enc_in_obj_ptrs: true
+  use_signed_tpos_enc_to_obj_ptrs: true
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False

custom_nodes/comfyui-segment-anything-2/sam2_configs/sam2.1_hiera_l.yaml

@@ -0,0 +1,120 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 144
+      num_heads: 2
+      stages: [2, 6, 36, 4]
+      global_att_blocks: [23, 33, 43]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+      window_spec: [8, 4, 16, 8]
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [1152, 576, 288, 144]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  no_obj_embed_spatial: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: true
+  proj_tpos_enc_in_obj_ptrs: true
+  use_signed_tpos_enc_to_obj_ptrs: true
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False

custom_nodes/comfyui-segment-anything-2/sam2_configs/sam2.1_hiera_s.yaml

@@ -0,0 +1,119 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 96
+      num_heads: 1
+      stages: [1, 2, 11, 2]
+      global_att_blocks: [7, 10, 13]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [768, 384, 192, 96]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  no_obj_embed_spatial: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: true
+  proj_tpos_enc_in_obj_ptrs: true
+  use_signed_tpos_enc_to_obj_ptrs: true
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False

custom_nodes/comfyui-segment-anything-2/sam2_configs/sam2.1_hiera_t.yaml

@@ -0,0 +1,121 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 96
+      num_heads: 1
+      stages: [1, 2, 7, 2]
+      global_att_blocks: [5, 7, 9]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [768, 384, 192, 96]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  # SAM decoder
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  no_obj_embed_spatial: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: true
+  proj_tpos_enc_in_obj_ptrs: true
+  use_signed_tpos_enc_to_obj_ptrs: true
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  # HieraT does not currently support compilation, should always be set to False
+  compile_image_encoder: False

custom_nodes/comfyui-segment-anything-2/sam2_configs/sam2_hiera_b+.yaml

@@ -0,0 +1,119 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 112
+      num_heads: 2
+      stages: [2, 3, 16, 3]
+      global_att_blocks: [12, 16, 20]
+      window_pos_embed_bkg_spatial_size: [14, 14]
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [896, 448, 224, 112]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  no_obj_embed_spatial: false
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  proj_tpos_enc_in_obj_ptrs: false
+  use_signed_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False

custom_nodes/comfyui-segment-anything-2/sam2_configs/sam2_hiera_l.yaml

@@ -0,0 +1,120 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 144
+      num_heads: 2
+      stages: [2, 6, 36, 4]
+      global_att_blocks: [23, 33, 43]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+      window_spec: [8, 4, 16, 8]
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [1152, 576, 288, 144]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  no_obj_embed_spatial: false
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  proj_tpos_enc_in_obj_ptrs: false
+  use_signed_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False

custom_nodes/comfyui-segment-anything-2/sam2_configs/sam2_hiera_s.yaml

@@ -0,0 +1,119 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 96
+      num_heads: 1
+      stages: [1, 2, 11, 2]
+      global_att_blocks: [7, 10, 13]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [768, 384, 192, 96]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  no_obj_embed_spatial: false
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  proj_tpos_enc_in_obj_ptrs: false
+  use_signed_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False

custom_nodes/comfyui-segment-anything-2/sam2_configs/sam2_hiera_t.yaml

@@ -0,0 +1,121 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 96
+      num_heads: 1
+      stages: [1, 2, 7, 2]
+      global_att_blocks: [5, 7, 9]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [768, 384, 192, 96]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  # SAM decoder
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  no_obj_embed_spatial: false
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  proj_tpos_enc_in_obj_ptrs: false
+  use_signed_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  # HieraT does not currently support compilation, should always be set to False
+  compile_image_encoder: False

custom_nodes/comfyui-tensorops/.gitattributes

@@ -0,0 +1,2 @@
+# Auto detect text files and perform LF normalization
+* text=auto

custom_nodes/comfyui-tensorops/.gitignore

@@ -0,0 +1,2 @@
+__pycache__/
+config_.py

custom_nodes/comfyui-tensorops/__init__.py

@@ -0,0 +1,3 @@
+from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
+
+__all__ = ['NODE_CLASS_MAPPINGS', 'NODE_DISPLAY_NAME_MAPPINGS']

custom_nodes/comfyui-tensorops/nodes/__init__.py

@@ -0,0 +1,54 @@
+from .channel_select import ChannelSelector
+from .mask_image import MaskImage
+from .save_surreal import SaveJsonToSurreal, SaveTextToSurreal
+from .fetch_surreal import FetchJsonFromSurreal
+from .foreground_mask import ForegroundMask
+from .save_to_s3 import SaveImageToS3
+from .redis import SaveToRedis, FetchFromRedis
+from .fal import FalDifferentialDiffusion, FalDiffusion
+from .background_select import BackgroundSelect
+from .layer_mask import GetLayerMask
+from .stream import SendImageOnWebSocket, SendJsonOnWebSocket
+from .separate_mask import SeparateMask
+from .face_swap import FaceSwap
+
+NODE_CLASS_MAPPINGS = {
+    "ChannelSelector": ChannelSelector,
+    "MaskImage": MaskImage,
+    "SaveImageToS3": SaveImageToS3,
+    "SaveJsonToSurreal": SaveJsonToSurreal,
+    "SaveTextToSurreal": SaveTextToSurreal,
+    "FetchJsonFromSurreal": FetchJsonFromSurreal,
+    "ForegroundMask": ForegroundMask,
+    "SaveToRedis": SaveToRedis,
+    "FetchFromRedis": FetchFromRedis,
+    "FalDifferentialDiffusion": FalDifferentialDiffusion,
+    "FalDiffusion": FalDiffusion,
+    "BackgroundSelect": BackgroundSelect,
+    "GetLayerMask": GetLayerMask,
+    "SendImageOnWebSocket": SendImageOnWebSocket,
+    "SendJsonOnWebSocket": SendJsonOnWebSocket,
+    "SeparateMask": SeparateMask,
+    "FaceSwap": FaceSwap
+}
+
+# A dictionary that contains the friendly/humanly readable titles for the nodes
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "ChannelSelector":"ChannelSelector",
+    "MaskImage": "MaskImage",
+    "SaveImageToS3": "SaveImageToS3",
+    "SaveJsonToSurreal": "SaveJsonToSurreal",
+    "SaveTextToSurreal": "SaveTextToSurreal",
+    "FetchJsonFromSurreal": "FetchJsonFromSurreal",
+    "ForegroundMask": "ForegroundMask",
+    "SaveToRedis": "SaveToRedis",
+    "FetchFromRedis": "FetchFromRedis",
+    "FalDifferentialDiffusion": "FalDifferentialDiffusion",
+    "FalDiffusion": "FalDiffusion",
+    "BackgroundSelect": "BackgroundSelect",
+    "GetLayerMask": "GetLayerMask",
+    "SendImageOnWebSocket": "SendImageOnWebSocket",
+    "SendJsonOnWebSocket": "SendJsonOnWebSocket",
+    "SeparateMask": "SeparateMask",
+    "FaceSwap": "FaceSwap"
+}

custom_nodes/comfyui-tensorops/nodes/background_select.py

@@ -0,0 +1,71 @@
+import torch
+
+
+def get_background_mask(tensor: torch.Tensor):
+    """
+    Function to identify the background mask from a batch of masks in a PyTorch tensor.
+
+    Args:
+        tensor (torch.Tensor): A tensor of shape (B, H, W, 1) where B is the batch size, H is the height, W is the width.
+
+    Returns:
+        List of masks as torch.Tensor and the background mask as torch.Tensor.
+    """
+    B, H, W = tensor.shape
+
+    # Compute areas of each mask
+    areas = tensor.sum(dim=(1, 2))  # Shape: (B,)
+    
+    # Find the mask with the largest area
+    largest_idx = torch.argmax(areas)
+    background_mask = tensor[largest_idx]
+
+    # Identify if the largest mask touches the borders
+    border_touched = (
+        torch.any(background_mask[0, :]) or
+        torch.any(background_mask[-1, :]) or
+        torch.any(background_mask[:, 0]) or
+        torch.any(background_mask[:, -1])
+    )
+
+    # If the largest mask doesn't touch the border, search for another one
+    if not border_touched:
+        for i in range(B):
+            if i != largest_idx:
+                mask = tensor[i]
+                border_touched = (
+                    torch.any(mask[0, :]) or
+                    torch.any(mask[-1, :]) or
+                    torch.any(mask[:, 0]) or
+                    torch.any(mask[:, -1])
+                )
+                if border_touched:
+                    background_mask = mask
+                    break
+    
+    # Reshape the masks to match the original tensor shape
+    return background_mask
+
+class BackgroundSelect:
+   
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "mask": ("MASK",),
+            },
+        }
+
+    RETURN_TYPES = ("MASK",)
+
+    FUNCTION = "main"
+
+    CATEGORY = "tensorops"
+
+    def main(self, mask: torch.Tensor):
+        # TODO loop through all masks
+        # identify the background mask
+        # return the background mask
+        background_mask = get_background_mask(mask)
+        return (background_mask,)
+