cstr/nafnet-sidd-GGUF

NAFNet-SIDD-width32 GGUF

GGUF conversion of megvii-research/NAFNet (MIT license) for use with CrispEmbed scan cleanup.

Model

NAFNet (Non-linear Activation Free Network) is a U-Net image restoration model that achieves state-of-the-art denoising without traditional nonlinear activations. Instead it uses SimpleGate (channel split + element-wise multiply) and Simplified Channel Attention.

• Architecture: U-Net with NAFBlocks
• Config: width=32, enc=[2,2,4,8], middle=12, dec=[2,2,2,2]
• Channels: 32 → 64 → 128 → 256 → 512 (middle) → 256 → 128 → 64 → 32
• Parameters: 29.2M
• Training: SIDD (Smartphone Image Denoising Dataset)
• Performance: PSNR 39.97 dB, SSIM 0.9599 on SIDD validation

Files

File	Type	Size	Notes
nafnet-sidd-w32-f16.gguf	F16	56 MB	Full precision weights
nafnet-sidd-w32-q8_0.gguf	Q8_0	30 MB	Recommended
nafnet-sidd-w32-q4_k.gguf	Q4_K	16 MB	Maximum compression

Usage with CrispEmbed

# CLI — preprocess scan before OCR
./build/crispembed --cleanup -m ocr_model.gguf --ocr scan.png

# Standalone cleanup
./build/crispembed --cleanup-only scan.png

from crispembed import CrispScanCleanup

cleanup = CrispScanCleanup("nafnet-sidd-w32-q8_0.gguf")
cleaned = cleanup.process("scan.png")  # numpy RGB array

Architecture Details

Each NAFBlock:

1. LayerNorm2d → Conv1x1 (c→2c) → DepthwiseConv3x3 → SimpleGate (2c→c)
2. Simplified Channel Attention (global avg pool → Conv1x1)
3. Conv1x1 (c→c) → residual connection (×beta)
4. LayerNorm2d → Conv1x1 (c→2c) → SimpleGate → Conv1x1 → residual (×gamma)

Downsampling: Conv2d stride 2, kernel 2×2. Upsampling: Conv1x1 + PixelShuffle(2).

License

MIT (megvii-research/NAFNet). The GGUF conversion does not change the license.

Citation

@article{chen2022simple,
  title={Simple Baselines for Image Restoration},
  author={Chen, Liangyu and Chu, Xiaojie and Zhang, Xiangyu and Sun, Jian},
  journal={arXiv preprint arXiv:2204.04676},
  year={2022}
}

Provenance and EU AI Act Art. 53 note

• Upstream model: megvii-research/NAFNet.
• Upstream licence: mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.