GFPGANv1.4-fp32

GFPGANv1.4-fp32 (MLX)

GFPGAN v1.4 blind face restoration converted to MLX NHWC safetensors for Apple Silicon. 87,143,276 parameters (params_ema), fp32.

• Architecture: GFPGANv1Clean — degradation-removal U-Net + StyleGAN2 (clean) decoder with channel-split SFT. 512×512 aligned face crops, RGB in [-1, 1].
• Layout: MLX NHWC. Conv (O,kH,kW,I); the 5-D modulated-conv kernels (1,O,kH,kW,I); every other 4-D tensor (biases, stored noise buffers, constant input) takes the same (0,2,3,1) transpose. Keys mirror the upstream state dict exactly.
• Noise: the checkpoint's stored noise buffers ship in the file (randomize_noise=False is the deterministic parity mode).
• dtype: fp32. fp16 is DISQUALIFIED — the e2e gate collapses (cosine ≈ −0.3, PSNR 8 dB); bf16 gates clean at 48 dB if you need it, but restoration ships fp32 here.

License

Apache-2.0. The upstream repo carries third-party carve-outs (NVIDIA StyleGAN2, DFDNet) that do NOT touch this artifact: the clean architecture contains no NVIDIA-derived code, and the v1.4 decoder prior was trained from scratch by BasicSR (StyleGAN2_512_Cmul1_FFHQ_B12G4_scratch_800k.pth, Apache-2.0), not from NVIDIA's FFHQ weights. Trained on FFHQ (dataset compilation CC-BY-NC-SA — the standard unsettled dataset-to-weights question that applies to every face model).

Consume

Swift (Apple Silicon): mlx-gfpgan-swift — GFPGANMLXCore.GFPGANv1Clean + MLXGFPGAN.GFPGANRestorePackage (MLXEngine imageRestore, Vision-based detect/align/paste).

Converted with oracle/convert.py in that repo; parity vs the PyTorch reference: key contract 285/285 tensors, per-stage taps 94/94 ≤ 5e-4 relative (fp32, CPU stream).