Upload README.md with huggingface_hub

README.md

@@ -17,13 +17,16 @@ language:
 
 # Moebius — ONNX (browser / WebGPU)
 
-ONNX exports of the [**Moebius**](https://huggingface.co/hustvl/Moebius) 0.22B lightweight
-image-inpainting model ([hustvl/Moebius](https://github.com/hustvl/Moebius), ECCV'26), packaged
-to run **fully client-side in a web browser** via
-[ONNX Runtime Web](https://onnxruntime.ai/docs/tutorials/web/) on the **WebGPU** backend.
+ONNX exports of the [Moebius](https://huggingface.co/hustvl/Moebius) image-inpainting model
+([hustvl/Moebius](https://github.com/hustvl/Moebius), ECCV'26; 0.22B parameters), for running in
+a web browser with [ONNX Runtime Web](https://onnxruntime.ai/docs/tutorials/web/) on the WebGPU
+backend.
 
-No text encoder, no tokenizer — Moebius conditions on a small learned embedding table, so the
-whole pipeline is three graphs plus a short DDIM loop you drive in JS.
+Moebius conditions on a learned embedding table rather than a text encoder, so there is no
+tokenizer or text model to export. The export is three graphs — VAE encoder, UNet, VAE decoder —
+and the sampling loop (DDIM with classifier-free guidance) runs in JavaScript.
+
+Live demo in your browser here: **[simonw.github.io/moebius-web/](https://simonw.github.io/moebius-web/)**. [Source code on GitHub](https://github.com/simonw/moebius-web).
 
 ## Files
 
@@ -52,10 +55,11 @@ that loads these files lives in the accompanying web demo.
 
 ## Precision
 
-These are **fp32** exports (chosen for guaranteed numeric parity with the reference pipeline).
-Parity vs PyTorch on CPU EP: decoder `max|Δ|≈5.7e-5`, unet `≈3.6e-6`. A full-pipeline check vs the
-Torch reference (identical noise) gives a decoded-image `mean|Δ|≈0.0022`. fp16 halves the download
-but is a quality gamble in the lambda layers and unstable for the VAE — validate before using.
+These are fp32 exports, for numeric parity with the reference pipeline. Parity vs PyTorch on the
+CPU execution provider: decoder `max|Δ|≈5.7e-5`, unet `≈3.6e-6`. A full-pipeline check against the
+PyTorch reference (identical initial noise) gives a decoded-image `mean|Δ|≈0.0022`. fp16 halves the
+download size, but can reduce quality in the lambda layers and is numerically unstable for this
+VAE; validate before use.
 
 ## License & attribution