MI-GAN 512 generator (ONNX, static 1×4×512×512)
Bare generator extracted from andraniksargsyan/migan
migan_pipeline_v2.onnx — upstream's bundled pre/post-processing (NonZero bbox
crop, blur blend, ScatterND paste-back) prevents CoreML from compiling the model,
so the wrapper is cut and reimplemented host-side.
Cut points: /Concat_23_output_0 → /model/synthesis/b512/Add_1_output_0.
Validated against upstream's own intermediate tensors: preprocessing bit-exact,
generator to float32 epsilon (2.4e-07).
Original work: MI-GAN (ICCV 2023), Picsart AI Research (PAIR). https://github.com/Picsart-AI-Research/MI-GAN — MIT License, Copyright (c) 2024 Picsart AI Research (PAIR). Trained on Places2.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support