imagingforgood/mi-gan-512-onnx

MI-GAN 512 (Places2), ONNX

An unchanged copy of the ONNX file of MI-GAN published by its first author, kept by Imaging for Good so that browser-perception can rely on it. We did not train, convert or modify this model.

Origin

• Model: MI-GAN by Picsart AI Research, MIT. Code: Picsart-AI-Research/MI-GAN. Paper: Sargsyan, Navasardyan, Xu and Shi, MI-GAN: A Simple Baseline for Image Inpainting on Mobile Devices, ICCV 2023 (paper).
• File: migan.onnx from andraniksargsyan/migan (Andranik Sargsyan, the first author) at commit 406830d0fa60666da0071c342ad2fbc8f30c5c64. The same repository has migan_pipeline_v2.onnx, which the MI-GAN README links; not copied here.

File

File	Bytes	SHA-256
migan.onnx	29,546,882	593eba0b7e04730f1b61c0a3cbca68d97d8d6a7ff5c6a44a7b9d7fcd880fc5ae

Contract

• image [1, 3, H, W] uint8 RGB and mask [1, 1, H, W] uint8: 255 = keep, 0 = fill. Any size; the model works at 512×512 inside.
• result [1, 3, H, W] uint8, the filled image (unmasked pixels from the input).

Training data

The model was trained on Places2, whose terms of use are for non-commercial research and education. The MIT licence covers the code and weights; check that the training data's terms fit your use.

Licence

MIT, for the code and the weights alike (MI-GAN's LICENSE-WEIGHTS); see LICENSE (copied from andraniksargsyan/migan, © 2024 Picsart AI Research).