RestoreFormerPlusPlus-fp32 (MLX)

RestoreFormer++ (TPAMI 2023) blind face restoration converted to MLX NHWC safetensors for Apple Silicon. 73,472,579 parameters (441 tensors), fp32.

• Architecture: VQVAEGANMultiHeadTransformer — VQ-GAN encoder/decoder over a 1024×256 ROHQD codebook, multi-scale multi-head cross-attention (enc attn @16, dec @[16, 32]). 512×512 aligned face crops, RGB in [-1, 1]. Fully deterministic.
• Source: the author's official RestoreFormer++.ckpt (v1.0.0 GitHub release); vqvae.* state re-exported through the instantiated architecture.
• Layout: MLX NHWC. Conv (O,kH,kW,I); GroupNorm vectors, biases, and the codebook embedding pass through. Keys mirror the upstream state dict (prefix stripped).
• dtype: fp32. Measured alternatives: fp16 50.1 dB vs the fp32 golden (viable), bf16 38.6 dB (mantissa-bound — fp16 beats bf16 here). Late-decoder activations reach ±14k.

License

Apache-2.0 — upstream is plain Apache-2.0 with no third-party carve-outs. Trained on FFHQ (dataset compilation CC-BY-NC-SA — the standard unsettled dataset-to-weights question).

Consume

Swift (Apple Silicon): mlx-restoreformer-swift — RestoreFormerMLXCore.RestoreFormer + MLXRestoreFormer.RestoreFormerRestorePackage (MLXEngine imageRestore, Vision-based detect/align/paste).

Parity vs the PyTorch reference: key contract 441/441 tensors; per-stage taps 40/40 ≤ 5e-4 relative (fp32, CPU stream); codebook indices 256/256 exact.