sam-bha commited on
Commit
c29f30d
·
verified ·
1 Parent(s): 6497862

Model card: clarify weights version (v0.0.5 = current), correct note on community port

Browse files
Files changed (1) hide show
  1. README.md +8 -5
README.md CHANGED
@@ -27,8 +27,10 @@ networks as hand-written WebGPU/WebGL2 shaders. These ONNX files let you use the
27
  anywhere ONNX Runtime runs: Python, native apps, servers, mobile, or `onnxruntime-web`.
28
 
29
  - **License:** MIT, code *and* weights. Training data provenance: [longpipe.dev/docs/licensing](https://longpipe.dev/docs/licensing).
30
- - **Weights:** identical to the SDK's production tiers (SDK v0.0.16), exported directly from the PyTorch
31
- training checkpoints. Every file is checked against PyTorch (max abs diff ≤ 7e-5, most ≤ 1e-6); see `onnx/manifest.json`.
 
 
32
 
33
  ## Quality and speed
34
 
@@ -129,9 +131,10 @@ not in these graphs. For a single-file temporal step, see the community port bel
129
  - Exported with ONNX opset 17 (`large`: opset 18). Tested with ONNX Runtime 1.23 (CPU).
130
  - Each `face` / `flow` / `matting` file contains its own copy of the tier's encoder. The SDK shares
131
  one encoder pass across all heads; if you need that, run the SDK.
132
- - Community port: [imbcmdth/longpipe-onnx](https://github.com/imbcmdth/longpipe-onnx) converted an
133
- earlier release (v0.0.5) by reconstructing the weights from the SDK's packed format, and includes a
134
- single-graph temporal step. Thanks! This repo is the official, current-weights export.
 
135
 
136
  ## Links
137
 
 
27
  anywhere ONNX Runtime runs: Python, native apps, servers, mobile, or `onnxruntime-web`.
28
 
29
  - **License:** MIT, code *and* weights. Training data provenance: [longpipe.dev/docs/licensing](https://longpipe.dev/docs/licensing).
30
+ - **Weights:** the SDK's current production weights (weights release `v0.0.5`, served at
31
+ `cdn.longpipe.dev/models/v/0.0.5/` and used by SDK 0.0.16), exported directly from the PyTorch training
32
+ checkpoints that produce those files. Every ONNX file is checked against PyTorch (max abs diff ≤ 7e-5,
33
+ most ≤ 1e-6); see `onnx/manifest.json`.
34
 
35
  ## Quality and speed
36
 
 
131
  - Exported with ONNX opset 17 (`large`: opset 18). Tested with ONNX Runtime 1.23 (CPU).
132
  - Each `face` / `flow` / `matting` file contains its own copy of the tier's encoder. The SDK shares
133
  one encoder pass across all heads; if you need that, run the SDK.
134
+ - Community port: [imbcmdth/longpipe-onnx](https://github.com/imbcmdth/longpipe-onnx) converted the same
135
+ weights by reconstructing them from the SDK's packed GPU format (xs/small/large/xl), and includes a
136
+ single-graph temporal step with the flow stabilizer. Thanks! This repo is the official export, made
137
+ directly from the training checkpoints, covering all five tiers in both orientations.
138
 
139
  ## Links
140