| --- |
| license: other |
| license_name: mixed-apache-2.0-bsd-3-clause-mit |
| license_link: LICENSE |
| library_name: onnx |
| pipeline_tag: image-to-image |
| tags: |
| - super-resolution |
| - jpeg-artifact-removal |
| - inpainting |
| - object-removal |
| - onnx |
| - directml |
| --- |
| |
| # Warpstead enhance models |
|
|
| ONNX files used by the Clean up and Upscale nodes and by the Remove brush of the image node of [Warpstead](https://warpstead.com). The app downloads each one the first time it's needed, checks its SHA-256 and runs it on the PC with ONNX Runtime, on the graphics card through DirectML when there is one. No photo leaves the computer. |
|
|
| | File | Model | Used by | License | |
| |---|---|---|---| |
| | `fbcnn-color.onnx` | [FBCNN](https://github.com/jiaxi-jiang/FBCNN), color | Clean up | Apache-2.0 | |
| | `realesrgan-x4plus.onnx` | [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN) x4plus, fp32 | Upscale, Photo | BSD-3-Clause | |
| | `realesr-general-x4v3-dn0.onnx`<br>`realesr-general-x4v3-dn50.onnx`<br>`realesr-general-x4v3-dn100.onnx` | Real-ESRGAN general x4v3, three denoise strengths | Upscale, Fast | BSD-3-Clause | |
| | `swinir-m-x4-classical.onnx` | [SwinIR](https://github.com/JingyunLiang/SwinIR)-M, classical x4, fixed 64×64 input | Upscale, Faithful | Apache-2.0 | |
| | `moebius-unet-dml.onnx` | [Moebius](https://github.com/hustvl/Moebius) inpainting UNet, rewritten for DirectML | Remove brush | Apache-2.0 | |
| | `moebius-vae-encoder.onnx`<br>`moebius-vae-decoder.onnx` | The VAE Moebius works with, from [PixelHacker](https://huggingface.co/hustvl/PixelHacker) | Remove brush | MIT | |
|
|
| The files are ONNX conversions of the official weights, made by the Warpstead developers. The three `realesr-general` files blend the two official realesr-general weights at a fixed denoise strength. The Moebius files start from the ONNX export by Simon Willison ([simonw/Moebius-ONNX](https://huggingface.co/simonw/Moebius-ONNX)); in the UNet the cross-attention Einsum nodes are rewritten as MatMul and the input shapes are fixed, so it runs fast on DirectML. The numbers it computes are the same. Copyright notices and full license texts are in [LICENSE](LICENSE). |
|
|
| Warpstead pins each download to a commit of this repository, so a file changed here never reaches a version that is already out. |
|
|
