ONNX image models for browser inference
This repository collects ONNX conversions of image restoration and upscaling models for local browser inference with ONNX Runtime Web and WebGPU. Most were converted with my pth2onnx-converter and are tested in onnx-web-upscale.
The collection includes upscalers, JPEG cleanup, denoise, deblur, exposure correction, and detail-recovery models. Performance and memory use depend on the browser, GPU, driver, source size, and selected model. Some exports include manual graph fixes or optimizations for ONNX Runtime Web compatibility.
Licensing provenance
Licensing provenance for some models in the repository.
| ONNX file | Original model and author | License |
|---|---|---|
deepwb_net_awb.onnx |
Deep White-Balance Editing by Mahmoud Afifi and Michael S. Brown | CC BY-NC-SA 4.0 |
1x-Film-Degrainer-1-000.onnx |
Film-Degrainer-1-000 by tika | CC0 1.0 |
1xDeJPG_SRFormer_light_fp32_64.onnx |
DeJPG_SRFormer_light by Helaman | CC BY 4.0 |
1xgaterv3_r_sharpen_fp16_op17_onnxslim.onnx |
1xgaterv3_r_sharpen by Philip Hofmann (unmodified release artifact) | CC BY 4.0 |
1xgaterv3_r_sharpen_fp32_op17.onnx |
1xgaterv3_r_sharpen by Philip Hofmann (unmodified release artifact) | CC BY 4.0 |
1xgaterv3_r_sharpen_fp32_op17_onnxslim.onnx |
1xgaterv3_r_sharpen by Philip Hofmann (unmodified release artifact) | CC BY 4.0 |
1x-Kim2091-DeJpeg-v0.onnx |
Kim2091 DeJPEG v0 by kim2091 | CC BY-NC-SA 4.0 |
1x-BW-Denoise.onnx |
BW Denoise by loganavter | CC BY-NC-SA 4.0 |
1xSkinContrast-SuperUltraCompact.onnx |
SkinContrast SuperUltraCompact by brigadeiro | CC0 1.0 |
1x-ITF-SkinDiffDetail-Lite-v1.onnx |
ITF SkinDiffDetail Lite v1 by intheflesh | CC BY-NC-SA 4.0 |
rife_v4.17_lite_v2.onnx |
Practical-RIFE 4.17-lite by Zhewei Huang, converted to ONNX by vs-mlrt | MIT |
rife_v4.22_lite_v2.onnx |
Practical-RIFE 4.22-lite by Zhewei Huang, converted to ONNX by vs-mlrt | MIT |
rife_v4.25_lite_v2.onnx |
Practical-RIFE 4.25-lite by Zhewei Huang, converted to ONNX by vs-mlrt | MIT |
rife_v4.25_v2.onnx |
Practical-RIFE 4.25 by Zhewei Huang, converted to ONNX by vs-mlrt | MIT |
Licensing
This is a mixed-license collection, so no single license applies to every file. Files listed as CC BY 4.0 require attribution and an indication of changes. Files listed as CC BY-NC-SA 4.0 additionally carry noncommercial and ShareAlike conditions. CC0 files remain available under CC0 1.0.
For models not listed in the table, check the original release and its license before reuse. The presence of a converted file in this repository does not by itself grant commercial-use rights.
rife_v4.25_lite_v2.onnx is the unmodified rife_v2 graph from the vs-mlrt
rife_v4.25_lite.7z release. The upstream Practical-RIFE README explicitly
places its linked trained-model content under MIT. Artifact SHA-256:
610b5de57cdcfbcce9914c23e60a1cd357779a6f9582a1bcfcb035f8eb38509b.
rife_v4.25_v2.onnx is the unmodified rife_v2 graph from the vs-mlrt
rife_v4.25.7z release. It has the same Practical-RIFE MIT model license and
vs-mlrt conversion provenance. Artifact SHA-256:
65c57a5e4abb17ad67faf35054291ac53affab0506d509d2e66698f6ecd75584.
rife_v4.17_lite_v2.onnx is the unmodified rife_v2 graph from the vs-mlrt
rife_v4.17_lite.7z release. It has the same Practical-RIFE MIT model license
and vs-mlrt conversion provenance. Artifact SHA-256:
4192e1db7db7d8a110a667b8776b9fe3d92deb1cce04676d5d57a5fd52d7578a.
rife_v4.22_lite_v2.onnx is the unmodified rife_v2 graph from the vs-mlrt
rife_v4.22_lite.7z release. It has the same Practical-RIFE MIT model license
and vs-mlrt conversion provenance. Artifact SHA-256:
2a1eca144923e6f5d9b10c07f8355883361ae378d67539e50bba1616effaa0cc.
Compatibility notes
The upstream 1xgaterv3_r_sharpen_fp16_op17_onnxslim.onnx release artifact is
mirrored for completeness, but ONNX Runtime 1.24 rejects its graph because an
internal Cast output is declared float16 where the graph expects float32.
Use 1xgaterv3_r_sharpen_fp32_op17_onnxslim.onnx for browser inference.
Credits
Original model weights, architectures, and training work belong to their respective authors, credited in their source releases. My contribution is ONNX conversion, packaging, and browser-compatibility work.