Obscura models
ONNX models downloaded on demand by Obscura.
NAFNet-SIDD-width32 (AI denoise)
| File | Precision | Size | SHA-256 |
|---|---|---|---|
nafnet-sidd-w32.onnx |
FP32 | 117,085,438 B | 092dc37894ee36f3ceef1359f73b0e566a24c9194138458cfd92b56d8f74647a |
nafnet-sidd-w32-fp16.onnx |
FP16 weights and maths, FP32 in/out | 58,957,563 B | 74bb136d6e78ebab8490c7db29b5bda37c270bd382a03990c66d9cc17271f581 |
- Model: NAFNet, "Simple Baselines for Image Restoration" (Chen, Chu, Zhang, Sun — ECCV 2022), https://github.com/megvii-research/NAFNet — MIT License, Copyright (c) 2022 megvii-model.
- Weights: Megvii's official
NAFNet-SIDD-width32.pth(trained on SIDD), SHA-25689c70e808d1783b6c07911306e106aaf0d4f7f3da8c61078b99ff7f8929a26f4, converted unchanged. - Export:
tools/ml/export_nafnet.pyin the Obscura repository (ONNX opset 17; FP16 by onnxconverter-common with FP32 inputs/outputs). - Input / output:
input→output, float32 NCHW RGB in 0..1, any height/width that are multiples of 16. The network's own padding is left out: the caller tiles.
MIT License text: https://github.com/megvii-research/NAFNet/blob/main/LICENSE