File size: 3,557 Bytes
58f2379 5653155 58f2379 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 | ---
library_name: onnxruntime
license: other
tags:
- webgpu
- onnx
- dinov2
- feature-visualization
---
# SAM_CAM_BAM Dream model pack
This is the immutable browser artifact pack for the experimental **Dream**
mode in [SAM_CAM_BAM](https://github.com/hammamiomar/SAM_CAM_BAM). It is not a
general-purpose pipeline. The app combines these files with a small handwritten
WebGPU block-transfer kernel.
Dream encodes a live source frame and an uploaded donor image with
DINOv2-B/14-with-registers, selects eight learned rank-3 block subspaces from a
clicked donor patch, transfers those projections to every source patch, and
decodes the edited patch field with an RAE ViT-B decoder. Inference stays in the
browser; this repository only serves static model bytes.
## Files
| File | Bytes | SHA-256 | Purpose |
| --- | ---: | --- | --- |
| `dinov2-vitb14-reg-q4f16.onnx` | 51,374,908 | `ef0feac82fd3fc1310aba8c38390fee86367b8c9e553a9846ce3488c29494ee5` | DINO encoder when shader-f16 is available |
| `dinov2-vitb14-reg-q4.onnx` | 59,978,957 | `02996745913d0654d8571cef0de689a7d3a4870b3deacf9b41d6c13b6dca7fd1` | Portable DINO encoder fallback |
| `dinov2-vitb14-reg-bsf-g8192-r3-f16.bin` | 38,547,520 | `54f6243b339af889ac826faf9fb5b535e3135e41068117ba6c78f95ad7ecbf61` | 8,192 orthonormal rank-3 frames, position mean, and encoder affine data |
| `rae-vitb-decoder-int8-qdq.onnx` | 87,946,239 | `cb31eec571791e3f3ab9cc03abd63841603ffbce85853423f47a75e53a907691` | Static 256-patch to 256×256 RGB decoder |
`manifest.json` records the complete source identities, tensor contracts, and
verification results. The pack is consumed by commit-pinned URLs; `main` is not
a runtime contract.
## Verification
The build script refuses source artifacts whose hashes differ from its pinned
inputs. The published pack passed:
- maximum BSF frame Gram error: `5.96e-7`;
- decoder int8-QDQ versus FP32 PSNR: `39.45 dB`;
- float16 versus float32 selector mean top-8 overlap: `1.0`; and
- exact ordered top-8 agreement on 254 of 256 fixture patches, including the
default strongest patch.
The reproducible build and checks live in
[`tools/dream_models.py`](https://github.com/hammamiomar/SAM_CAM_BAM/blob/main/tools/dream_models.py).
## Provenance and terms
- The two encoder graphs are byte-identical copies from
[`onnx-community/dinov2-with-registers-base`](https://huggingface.co/onnx-community/dinov2-with-registers-base),
revision `fcd050203baf48e192055356f379b5949a111255`. DINOv2 code and model
weights are Apache-2.0.
- The decoder was exported from RAE source
[`bytetriper/RAE`](https://github.com/bytetriper/RAE) at
`a4d18c4db766419cbe7cb8c02cd9f7ceb0ec9041`, then its 74 constant matrix
weights were quantized to int8 QDQ. The RAE source repository is MIT.
- The ViT-B decoder checkpoint is
`decoders/dinov2/wReg_base/decB_ganv3/dinov2_decoder.pt` from
[`nyu-visionx/RAE-collections`](https://huggingface.co/nyu-visionx/RAE-collections),
revision `1be4f03273523431f099a934da4cf1940dc6039f`. That repository declares
the MIT license.
- The BSF field is an Omar Hammami-trained artifact over calibrated DINOv2
patch states. Its exact checkpoint and calibration hashes are in the
manifest.
## Limits
Dream is a creative feature-visualization experiment, not segmentation,
depth, restoration, or identity-preserving image editing. Inputs are resampled
to 224×224 and output is 256×256. Results can distort content, faces, text, and
fine structure and inherit limitations from the source models and their
training data.
|