Denoising on SSID (test)
50PSNRModular Neural ISP
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Modular Neural ISPTraining Data=Generic (Wide range of noise profiles), Model Variant=base2025.12 | 50 | 97.8 | |
| Modular Neural ISPTraining Data=Generic (Wide range of noise profiles), Model Variant=large2025.12 | 49.55 | 97.9 | |
| Modular Neural ISPTraining Data=Generic (Wide range of noise profiles), Model Variant=lite2025.12 | 49.09 | 97.1 | |
| Modular Neural ISPTraining Data=MIT Adobe 5K [26], Model Variant=base2025.12 | 48.28 | 96.7 | |
| Modular Neural ISPTraining Data=MIT Adobe 5K [26], Model Variant=large2025.12 | 48.12 | 97.1 | |
| Modular Neural ISPTraining Data=MIT Adobe 5K [26], Model Variant=lite2025.12 | 48.05 | 96.6 | |
| Modular Neural ISPTraining Data=S24 [16], Model Variant=large2025.12 | 47.47 | 95.6 | |
| Modular Neural ISPTraining Data=S24 [16], Model Variant=base2025.12 | 47.28 | 95.2 | |
| Modular Neural ISPTraining Data=S24 [16], Model Variant=lite2025.12 | 46.53 | 94 | |
| RestormerArchitecture=Restormer2024.06 | 40.02 | 97.06 | |
| NAFNetArchitecture=NAFNet2024.06 | 39.97 | 95.99 | |
| Restormer + FPArchitecture=Restormer, Pooling=FrequencyPreservedPooling2024.06 | 39.88 | 97.01 | |
| Restormer + FP + DropHighArchitecture=Restormer, Pooling=FrequencyPreservedPooling, DropHigh=true2024.06 | 39.88 | 97.91 | |
| Restormer + FP + FirstLayerDropArchitecture=Restormer, Pooling=FrequencyPreservedPooling, FirstLayerDrop=true2024.06 | 39.88 | 97.01 | |
| Restormer + FreqAvgUpArchitecture=Restormer, Upsampling=FreqAvgUp2024.06 | 39.86 | 97.01 | |
| Restormer + FP + FirstLayerDrop + FreqAvgUpArchitecture=Restormer, Pooling=FrequencyPreservedPooling, FirstLayerDrop=true, Upsampling=FreqAvgUp2024.06 | 39.86 | 97.01 | |
| Restormer + FP + DropHigh + FreqAvgUp (BoA)Architecture=Restormer, Pooling=FrequencyPreservedPooling, DropHigh=true, Upsampling=FreqAvgUp, Protocol=BoA2024.06 | 39.85 | 97 | |
| Restormer + FP + FreqAvgUp (BoA)Architecture=Restormer, Pooling=FrequencyPreservedPooling, Upsampling=FreqAvgUp, Protocol=BoA2024.06 | 39.85 | 97.01 | |
| NAFNet + FreqAvgUpArchitecture=NAFNet, Upsampling=FreqAvgUp2024.06 | 39.75 | 95.85 | |
| NAFNet + FPArchitecture=NAFNet, Pooling=FrequencyPreservedPooling2024.06 | 39.73 | 95.84 | |
| NAFNet + FP + FirstLayerDrop + FreqAvgUpArchitecture=NAFNet, Pooling=FrequencyPreservedPooling, FirstLayerDrop=true, Upsampling=FreqAvgUp2024.06 | 39.72 | 95.82 | |
| NAFNet + FP + FirstLayerDropArchitecture=NAFNet, Pooling=FrequencyPreservedPooling, FirstLayerDrop=true2024.06 | 39.71 | 95.83 | |
| NAFNet + FP + FreqAvgUp (BoA)Architecture=NAFNet, Pooling=FrequencyPreservedPooling, Upsampling=FreqAvgUp, Protocol=BoA2024.06 | 39.64 | 95.77 | |
| NAFNet + FP + DropHigh + FreqAvgUpArchitecture=NAFNet, Pooling=FrequencyPreservedPooling, DropHigh=true, Upsampling=FreqAvgUp2024.06 | 39.54 | 95.77 | |
| NAFNet + FP + DropHighArchitecture=NAFNet, Pooling=FrequencyPreservedPooling, DropHigh=true2024.06 | 36.57 | 95.74 |