MI-GAN (ONNX) — object-removal inpainting

migan_pipeline_v2.onnx — an ONNX export of MI-GAN for image inpainting (erasing a masked object and filling the hole). This "pipeline" variant takes a full-size image + mask and returns the composited result, so it runs client-side with onnxruntime-web on the WASM or WebGPU execution providers with no pre/post resize step.

A verbatim mirror of andraniksargsyan/migan's migan_pipeline_v2.onnx, hosted so a single, immutable, CORS-enabled copy backs the Edge Tools "Remove Object" tool.

Inputs / outputs

• Inputs: image uint8 [1, 3, H, W], mask uint8 [1, 1, H, W]
• Output: result uint8 [1, 3, H, W] — the inpainted image, already composited at the input resolution
• RGB, channels-first (CHW). H and W are dynamic (the graph resizes internally). Mask polarity: 0 = erase (the hole), 255 = keep.

Provenance & license

• Algorithm / weights: MI-GAN — MI-GAN: A Simple Baseline for Image Inpainting on Mobile Devices (Sargsyan et al., ICCV 2023), Picsart-AI-Research/MI-GAN. Code and weights are released under the MIT license.
• ONNX export: andraniksargsyan/migan (migan_pipeline_v2.onnx).
• The upstream MIT LICENSE (© 2024 Picsart AI Research) is included in this repo.