In-browser image inpainting with the Moebius model (0.22B parameters, ECCV'26), running client-side via ONNX Runtime Web on the WebGPU backend.

• Live demo: https://simonw.github.io/moebius-web/
• Model weights (ONNX): https://huggingface.co/simonw/Moebius-ONNX

Ported to ONNX by Claude Opus 4.8. Full Claude Code transcript.

Paint over a region of an image to replace it; the denoising loop runs locally on your GPU. The first run downloads ~1.27 GB of weights from Hugging Face (then browser-cached). A WebGPU-capable browser (recent Chrome or Safari) is required.

Moebius conditions on a learned embedding table rather than a text encoder, so there is no tokenizer or text model. The export is three ONNX graphs — VAE encoder, UNet, VAE decoder — and the sampling loop (DDIM with classifier-free guidance, 9-channel latent assembly, pre/post-processing) runs in TypeScript. See web/src/pipeline.ts.

Notes:

• Fixed 512×512 resolution (the model's cross-attention uses a position embedding tied to the trained resolution); non-square inputs are letterboxed.
• VAE scaling_factor = 0.13025 (a custom VAE, not the usual SD 0.18215).
• The self-attention pos_conv is exported as Conv2d rather than Conv3d so the graph compiles on Safari's Metal WebGPU backend.

cd web
npm install
# point the app at local model files, or set VITE_MODEL_BASE to the HF repo:
npm run dev

The model files (unet.onnx, vae_encoder.onnx, vae_decoder.onnx) are not in this repo — they live in the Hugging Face model repo. For local dev, symlink them into web/public/models/, or build with VITE_MODEL_BASE=https://huggingface.co/simonw/Moebius-ONNX/resolve/main.

Pushing to main triggers .github/workflows/deploy.yml, which builds web/ (with the GitHub Pages base path and the Hugging Face model base) and deploys to GitHub Pages.

Apache 2.0, inherited from the upstream hustvl/Moebius. The model is by Duan, Xu, et al.; this repository is a browser port and ONNX conversion.