BryanHaley/photowagon-models

Inpainting models for Core ML

Models for PhotoWagon's AI Remove.

MI-GAN (migan/, in use)

• migan/MIGAN.mlpackage.zip: MI-GAN 512 Places2 (Sargsyan et al., Picsart AI Research), converted to Core ML with migan/convert.py from migan_512_places2.pt (SHA-256 1d6087eee0aac8923ad2606be5d8caeb4824d3e4de331995e420c74e124a466a). Zip SHA-256 c279edffd937627477b716376cea8e51aa8ac42c1691c8b47365020b915903c8.
• Inputs: image (RGB, 512 × 512) and mask (grayscale, 512 × 512, white = the hole). Output: output (RGB, 512 × 512). iOS 17 or later.
• Licence: MIT for code and weights (migan/LICENSE). Trained on Places2, whose images carry research-use terms.

LaMa (lama/, no longer used)

A mirror of the Core ML conversion of LaMa (Resolution-robust Large Mask Inpainting with Fourier Convolutions, Suvorov et al.), for PhotoWagon's AI Remove.

• lama/LaMa.mlpackage.zip: unchanged from mlboydaisuke/coreml-zoo (revision a3d12de43e1b6131cd05cdd9027e0c4978d436e5), converted by Daisuke Majima (john-rocky). SHA-256 b57b8451a1a86c00aea52d75230fb5f49d3076eec67403192758c9d2b59c0e69.
• Inputs: image (RGB, 800 × 800) and mask (grayscale, 800 × 800, white = the hole). Output: output (RGB, 800 × 800). iOS 17 or later, compute units: all.
• Licence: Apache-2.0 (LICENSE, from the LaMa repository). The weights were trained on Places2, whose images carry research-use terms.