Researcher ports 0.2B Moebius image inpainting model to run in-browser via WebGPU

Simon Willison reports successfully running the 0.2B-parameter Moebius image inpainting model in a web browser using WebGPU, after converting the PyTorch model to ONNX with Claude Code. The original Moebius release required PyTorch and NVIDIA CUDA, but the converted version runs client-side in the browser.

The process involved using ONNX Runtime Web with a WebGPU backend, which sits below libraries like Transformers.js. The author did not write any code; instead, they guided Claude Code through the conversion and deployment steps, including publishing the 1.24GB ONNX weights to Hugging Face and hosting the frontend on GitHub Pages.

The resulting demo is available at simonw.github.io/moebius-web and allows users to upload an image, select regions to remove, and run the inpainting model entirely in the browser. The first load downloads approximately 1.3GB of model weights, but subsequent reloads leverage browser caching to avoid re-downloading.

Willison notes this as an example of "vibe coding," where the primary contributions were testing, suggesting small improvements, and pointing the agent toward examples of desired functionality. The author also documented the process and learned that Chrome, Firefox, and Safari can all run this class of model in the browser.