CardAnvilModels/Moebius-pretrained-ONNX

Moebius (pretrained) — ONNX for WebGPU

ONNX exports of Moebius (hustvl, ECCV'26), a 0.22B-parameter latent diffusion model for removing things from images, for in-browser inference with ONNX Runtime Web on WebGPU. Used by Card Anvil's watermark remover.

This repository carries the general pretrained checkpoint — Moebius before it was fine-tuned on any one benchmark — rather than the Places2 fine-tune that simonw/Moebius-ONNX exports. The pretrained model was distilled from PixelHacker over its 14M image–mask pairs; the fine-tunes narrow it to one benchmark's photos.

Files

File	Graph	Input → output	Size
unet.onnx	Student denoiser, pretrained checkpoint	latent (B,9,64,64), timesteps (B,), input_ids (B,10) → noise (B,4,64,64)	907 MB
vae_encoder.onnx	VAE encoder	image (B,3,512,512) → moments (B,8,64,64)	137 MB
vae_decoder.onnx	VAE decoder	latent (B,4,64,64) → image (B,3,512,512)	198 MB

The UNet's inputs and outputs are identical to simonw/Moebius-ONNX's, so it drops into any pipeline written for that export. The two VAE graphs are copied unchanged from simonw/Moebius-ONNX at 5bf1ef5 (the VAE is PixelHacker's); all four Moebius checkpoints share it.

How the UNet was made

export_unet.py (here) is simonw's reexport_unet_no_conv3d.py pointed at pretrained/diffusion_pytorch_model.bin from hustvl/Moebius at cd01f47. Each self-attention pos_conv Conv3d is replaced by its equivalent Conv2d so the graph compiles on every WebGPU backend, then the model is exported with opset 18 at a static 512 × 512 (64 × 64 latent) — its cross-attention's position embedding is tied to that size.

Run on CPU (PyTorch 2.7.1), with upstream's CUDA-only teacher import (fla) skipped:

• Conv3d → Conv2d swap: max |Δ| 1.8e-6 against the unpatched model
• ONNX Runtime (CPU) against PyTorch: max |Δ| 3.1e-6
• 0 three-dimensional convolutions left in the graph

Using it

The sampling loop runs outside the graphs; it must match Moebius's:

• VAE scaling factor 0.13025 (not Stable Diffusion's 0.18215): encode as mean(moments[:, :4]) × 0.13025, decode latent / 0.13025.
• UNet input: concat([noisy latent (4), mask (1), masked image's latent (4)]).
• input_ids: rows [0…9] conditioned, [10…19] unconditioned, for classifier-free guidance (2 to 2.5 per the paper).
• DDIM: scaled_linear betas from 0.00085 to 0.012, 1000 training steps, no clipping; 20 steps at strength 0.99 run 19.

Licence

Apache-2.0, as is Moebius's code and weights and simonw/Moebius-ONNX. (hustvl's Hugging Face cards for Moebius and PixelHacker say MIT; their GitHub repositories say Apache-2.0.) The files here are converted from those works: the UNet re-exported as described above, the VAE graphs unchanged.

Citation

@inproceedings{DuanAndXu2026Moebius,
  title     = {Moebius: 0.2B Lightweight Image Inpainting Framework with 10B-Level Performance},
  author    = {Kangsheng Duan and Ziyang Xu and Wenyu Liu and Xiaohu Ruan and Xiaoxin Chen and Xinggang Wang},
  year      = {2026},
  booktitle = {European Conference on Computer Vision (ECCV)}
}