Jannchie/moebius-inpaint-onnx

Moebius inpainting — ONNX

ONNX export of Moebius (hustvl, ft_places2 weights) together with the PixelHacker VAE it uses. It runs on onnxruntime with no PyTorch or triton, and is the object-removal model of LLR.

All three graphs are fp32 at a fixed 512 × 512 (the training resolution). Pad or crop each region to a square before you run them.

file	inputs	output
moebius_unet.onnx	x [1,9,64,64] = noisy latent ‖ mask ‖ masked latent, t [1] int64, guidance scalar	eps [1,4,64,64]
vae_encoder.onnx	image [1,3,512,512] in [-1, 1]	scaled latent mean [1,4,64,64]
vae_decoder.onnx	latent [1,4,64,64]	image [1,3,512,512] in [0, 1]

Classifier-free guidance happens inside the UNet: both branches run as one batch of two. The conditioning is a constant table of learned embeddings.

Sampling

• Sampler: DDIM, scaled_linear betas from 0.00085 to 0.012, leading timestep spacing, eta 0.
• Settings: 20 steps, guidance 2.0.
• Start from pure noise (strength 1). The upstream pipeline defaults to strength 0.99, which lets the removed object leak back into the fill.
• Mask: 1 marks the region to fill. Downsample it to 64 × 64 for the UNet input.

A reference implementation in numpy is LLR's apps/worker/src/llr_worker/erase.py. Speed on an RTX 3090 with onnxruntime-directml: about 95 ms per UNet step.

The export script is apps/worker/scripts/export_moebius_onnx.py in the LLR repo, run on Moebius commit b88d462.

License

MIT, following the upstream weights (hustvl/Moebius, hustvl/PixelHacker). Credit goes to the Moebius authors.