How megvii-research/nafnet works
NAFNet is a PyTorch image-restoration toolbox for denoising, deblurring, and stereo super-resolution. The main user workflow is to provide degraded image(s), run an inference entry point, and save restored output. Training adds paired-image data preparation, model optimization, and validation metrics. The graph also captures reusable architecture, data, and runtime-support layers. Relationships are limited to those stated in the README or supported by the supplied source index; unsampled implementation details are not inferred.
Main entry points: train.py Training loop · image_restoration_model.py Restoration task · data Image and video datasets · NAFNet_arch.py NAFNet
Components 23
User workflows
- Image inference
basicsr/demo.py - Stereo inference
basicsr/demo_ssr.py - Prediction entry
predict.py
Data pipeline
- Image and video datasets
basicsr/data - Dataset utilities
basicsr/data/data_util.py - Image transforms
basicsr/data/transforms.py - Data prefetch
basicsr/data/prefetch_dataloader.py
Restoration models
- Restoration task
basicsr/models/image_restoration_model.py - NAFNet
basicsr/models/archs/NAFNet_arch.py - Baseline network
basicsr/models/archs/Baseline_arch.py - NAFSSR
basicsr/models/archs/NAFSSR_arch.py - Architecture helpers
basicsr/models/archs/arch_util.py - Local architecture
basicsr/models/archs/local_arch.py - Training losses
basicsr/models/losses/losses.py
Training and evaluation
- Training loop
basicsr/train.py - Model base
basicsr/models/base_model.py - Quality metrics
basicsr/metrics - Option parsing
basicsr/utils/options.py - Storage access
basicsr/utils/file_client.py - Distributed runtime
basicsr/utils/dist_util.py - Logging
basicsr/utils/logger.py - Image operations
basicsr/utils/matlab_functions.py
Other components
- Image user
Connections 24
- Image user → Image inference: runs inference
- Image user → Stereo inference: runs stereo
- Image user → Prediction entry: runs prediction
- Image inference → Restoration task: restores images
- Stereo inference → NAFSSR: runs stereo SR
- Prediction entry → Restoration task: creates model
- Training loop → Option parsing: parses options
basicsr/train.py - Training loop → Image and video datasets: loads data
basicsr/train.py - Training loop → Data prefetch: uses prefetch
basicsr/train.py - Training loop → Restoration task: creates model
- Training loop → Distributed runtime: initializes distributed
basicsr/train.py - Training loop → Logging: sets up logging
- Training loop → Quality metrics: evaluates output
- Image and video datasets → Dataset utilities: uses utilities
basicsr/data/video_test_dataset.py - Image and video datasets → Image transforms: uses transforms
basicsr/data/data_util.py - Restoration task → Model base: inherits from
basicsr/models/image_restoration_model.py - Restoration task → NAFNet: builds network
- Restoration task → Training losses: configures losses
basicsr/models/image_restoration_model.py - NAFNet → Architecture helpers: uses helpers
basicsr/models/archs/NAFNet_arch.py - NAFNet → Local architecture: uses modules
basicsr/models/archs/NAFNet_arch.py - Baseline network → Architecture helpers: uses helpers
basicsr/models/archs/Baseline_arch.py - Baseline network → Local architecture: uses modules
basicsr/models/archs/Baseline_arch.py - Quality metrics → Image operations: uses image ops
basicsr/metrics/metric_util.py - Image and video datasets → Storage access: reads files