the weight blobs moved to imbcmdth/longpipe-weights; README as before

README.md

@@ -1,65 +1,38 @@
-# longpipe-onnx
-
-Longpipe matting models for ffrwd model pins. The network and weights
-are [sb2702/longpipe](https://github.com/sb2702/longpipe), MIT licensed
-(code and weights; see their licensing page for dataset provenance).
-This repo holds two kinds of file: longpipe's own weight blobs, as the
-project serves them, and faithful ONNX conversions of them.
-
-## Weight blobs
-
-The files longpipe's own WebGPU code loads, byte for byte as served at
-cdn.longpipe.dev/models/v/0.0.5/ (its manifest.json carries the same
-sizes and digests). One per tier in f32, and the f16 packing of each.
-ffrwd/longpipe 0.3 and later pin these and run longpipe's code on the
-sidecar's GPU; the `tier` parameter picks one.
-
-| file | tier | bytes | sha256 |
-|---|---|---|---|
-| model_xs.bin | xs | 4300564 | 501cd8fc26a260552c62d22cde4d5ab7b32087921f42ba2539a9b5300d3f8c9d |
-| model_small.bin | small | 4232672 | 56430de7da6cfd6b834753f912441fbca00ad397790e308b8dc041e5c0b03394 |
-| model_medium.bin | medium | 14074452 | d6684cecaa1cf0570626872b89eef3690e41e3b861fc67f8916732b549d50945 |
-| model_large.bin | large | 14075812 | 5294fbbfd1eac70a6ee8d99dbe4de9a6d8c459d2c8a893c24402b35464e121f5 |
-| model_xl.bin | xl | 30463184 | c1ea11f3ff2a17eb60fe53c5d37d91c410f6d992ecb245236b21c7ef80217032 |
-| model_xs.f16.bin | xs | 2153062 | f1786bf9ffb03916fd265c253f4928ec4a7576855162aacc856aaf90908899e9 |
-| model_small.f16.bin | small | 2119038 | 8fea35e96ee8e37d1b2b6e901a2a7f8df1ce2ec14e1a1227636df6b410e2c11f |
-| model_medium.f16.bin | medium | 7040874 | 7d8bf5d787da24e679193e70db1e7b9ca62f473ef79d8f14db1604dda8d99dd5 |
-| model_large.f16.bin | large | 7041646 | 40ad8b9f538cf70b89b09a41f916ab441610c9a92ff4222a27071fa22fe92ae9 |
-| model_xl.f16.bin | xl | 15236450 | 4f25cb0c267d8966329ea7882afd42a64856786d5de1a276b67eb20f42043fe4 |
-
-Each blob is a u32 little-endian header length, a JSON index of the
-tensors, then the packed weights in the layouts longpipe's shaders read.
-
-## ONNX conversions
-
-Used by ffrwd/longpipe 0.2, which ran the xl system through ONNX Runtime.
-
-| file | preset | input | sha256 |
-|---|---|---|---|
-| matting_xl_temporal_f32.onnx | xl system | see notes | 67f79a367c4014d9df631afb002c99bf6efb1aa4ae122df175b45f2f92d611b6 |
-| matting_xl_refined_f32.onnx | xl system | see notes | 54b73fcc90335f4ef5274039f7ae09bf5462b2091172695c399bd8078b6b4a69 |
-| flow_xl_f32.onnx | xl system | see notes | e39ee98cb8ecff36e416ad234bb90b5301faa302becf4e82fec61869aa218df3 |
-| face_xl_f32.onnx | xl system | see notes | fb3d2ef4f6de97cc51708ed56ec9fb6e9fff8a5a0daeebf952a0f7ef8905cbb3 |
-| landmark_mesh_f32.onnx | xl system | see notes | e930cc43f06288ba53ba2a3251e2fd3c42995a5c39324c56988648cc32e62edc |
-| matting_large_f32.onnx | large | 1x3x160x256 RGB [0,1] | 5007ee2a9fe53dbdebfd2d6c9baba270199aaee9ad1189df9227030b51f21b70 |
-| matting_xl_f32.onnx | xl | 1x3x192x320 RGB [0,1] | 996cf2779c5b7c690b35ee1ef5e8890f8d2497e12dea97dc13ddd18e4132d04c |
-| matting_small_f32.onnx | small | 1x3x112x192 RGB [0,1] | d5a8c23461c253366ef9b87c59448e3d8c93062a65e25db6591b8def997f6afc |
-
-Converted from the f32 blobs above by reconstruction in PyTorch: the
-blob's own JSON index locates each tensor, the WGSL mat4x4 packing
-inverts to dense OIHW (the inverse of longpipe's own
-scripts/generate_fixtures.py packer), and the export holds
-PyTorch-to-ONNX parity under 1e-4. Conversion scripts live with the
-ffrwd project.
-
-The xl system files: `matting_xl_refined_f32.onnx` is base+wrapper
-in one graph (frame in, refined alpha out, canvas 1280x768);
-`matting_xl_temporal_f32.onnx` is the whole per-frame production
-step with the optical-flow stabilizer, state as explicit inputs and
-outputs (previous frame, four encoder taps, alpha and envelope);
-`flow_xl_f32.onnx` is the flow net alone; `face_xl_f32.onnx` emits
-five facial heatmaps at base/8; `landmark_mesh_f32.onnx` regresses
-478 landmarks from a 256x256 ImageNet-normalized crop (its blob is
-their separate model_landmark_mesh.bin, sha ca534744...3b87).
-
-MIT, matching upstream.
+# longpipe-onnx
+
+Longpipe matting models converted to ONNX for use as ffrwd model
+pins. The network and weights are [sb2702/longpipe](https://github.com/sb2702/longpipe),
+MIT licensed (code and weights; see their licensing page for dataset
+provenance). This repo holds faithful conversions, nothing more.
+
+| file | preset | input | sha256 |
+|---|---|---|---|
+| matting_xl_temporal_f32.onnx | xl system | see notes | 67f79a367c4014d9df631afb002c99bf6efb1aa4ae122df175b45f2f92d611b6 |
+| matting_xl_refined_f32.onnx | xl system | see notes | 54b73fcc90335f4ef5274039f7ae09bf5462b2091172695c399bd8078b6b4a69 |
+| flow_xl_f32.onnx | xl system | see notes | e39ee98cb8ecff36e416ad234bb90b5301faa302becf4e82fec61869aa218df3 |
+| face_xl_f32.onnx | xl system | see notes | fb3d2ef4f6de97cc51708ed56ec9fb6e9fff8a5a0daeebf952a0f7ef8905cbb3 |
+| landmark_mesh_f32.onnx | xl system | see notes | e930cc43f06288ba53ba2a3251e2fd3c42995a5c39324c56988648cc32e62edc |
+| matting_large_f32.onnx | large | 1x3x160x256 RGB [0,1] | 5007ee2a9fe53dbdebfd2d6c9baba270199aaee9ad1189df9227030b51f21b70 |
+| matting_xl_f32.onnx | xl | 1x3x192x320 RGB [0,1] | 996cf2779c5b7c690b35ee1ef5e8890f8d2497e12dea97dc13ddd18e4132d04c |
+| matting_small_f32.onnx | small | 1x3x112x192 RGB [0,1] | d5a8c23461c253366ef9b87c59448e3d8c93062a65e25db6591b8def997f6afc |
+
+Converted from the f32 weight blobs served at
+cdn.longpipe.dev/models/v/0.0.5/ (small blob 56430de7...b03394,
+large 5294fbbf...4e121f5, xl c1ea11f3...80217032) by
+reconstruction in PyTorch: the blob's own JSON index locates each
+tensor, the WGSL mat4x4 packing inverts to dense OIHW (the inverse
+of longpipe's own scripts/generate_fixtures.py packer), and the
+export holds PyTorch-to-ONNX parity under 1e-4. Conversion scripts
+live with the ffrwd project.
+
+The xl system files: `matting_xl_refined_f32.onnx` is base+wrapper
+in one graph (frame in, refined alpha out, canvas 1280x768);
+`matting_xl_temporal_f32.onnx` is the whole per-frame production
+step with the optical-flow stabilizer, state as explicit inputs and
+outputs (previous frame, four encoder taps, alpha and envelope);
+`flow_xl_f32.onnx` is the flow net alone; `face_xl_f32.onnx` emits
+five facial heatmaps at base/8; `landmark_mesh_f32.onnx` regresses
+478 landmarks from a 256x256 ImageNet-normalized crop (its blob is
+their separate model_landmark_mesh.bin, sha ca534744...3b87).
+
+MIT, matching upstream.

model_large.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:5294fbbfd1eac70a6ee8d99dbe4de9a6d8c459d2c8a893c24402b35464e121f5
-size 14075812

model_large.f16.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:40ad8b9f538cf70b89b09a41f916ab441610c9a92ff4222a27071fa22fe92ae9
-size 7041646

model_medium.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d6684cecaa1cf0570626872b89eef3690e41e3b861fc67f8916732b549d50945
-size 14074452

model_medium.f16.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:7d8bf5d787da24e679193e70db1e7b9ca62f473ef79d8f14db1604dda8d99dd5
-size 7040874

model_small.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:56430de7da6cfd6b834753f912441fbca00ad397790e308b8dc041e5c0b03394
-size 4232672

model_small.f16.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8fea35e96ee8e37d1b2b6e901a2a7f8df1ce2ec14e1a1227636df6b410e2c11f
-size 2119038

model_xl.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:c1ea11f3ff2a17eb60fe53c5d37d91c410f6d992ecb245236b21c7ef80217032
-size 30463184

model_xl.f16.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:4f25cb0c267d8966329ea7882afd42a64856786d5de1a276b67eb20f42043fe4
-size 15236450

model_xs.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:501cd8fc26a260552c62d22cde4d5ab7b32087921f42ba2539a9b5300d3f8c9d
-size 4300564

model_xs.f16.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:f1786bf9ffb03916fd265c253f4928ec4a7576855162aacc856aaf90908899e9
-size 2153062