How xinntao/real-esrgan works
Real-ESRGAN is an image/video restoration and super-resolution system. The main user workflow is: select media and a model, load weights, run tiled or whole-image enhancement, optionally enhance faces, and write restored images or encoded video. A second major workflow trains models from synthetic degradations or paired data using BasicSR. The graph includes documented anime/general models, video processing, arbitrary scaling, and optional GFPGAN integration. Some orchestration is inferred from sampled entrypoints and README capabilities; unsampled BasicSR internals are left external.
Main entry points: utils.py RealESRGAN Upsampler · inference_realesrgan.py Image Inference · inference_realesrgan_video.py Video Inference · realesrgan_dataset.py Synthetic Dataset
Components 20
Inference Entry Points
- Image Inference
inference_realesrgan.py - Video Inference
inference_realesrgan_video.py
Restoration Runtime
- RealESRGAN Upsampler
realesrgan/utils.py - Tile Processing
realesrgan/utils.py - SRVGG Network
realesrgan/archs/srvgg_arch.py - BasicSR Networks
- Model Weights
weights/README.md - Face Enhancer
Video Pipeline
- Video Reader
inference_realesrgan_video.py - Video Writer
inference_realesrgan_video.py
Training System
- Training Entry
realesrgan/train.py - Synthetic Dataset
realesrgan/data/realesrgan_dataset.py - Paired Dataset
realesrgan/data/realesrgan_paired_dataset.py - Degradation Model
realesrgan/models/realesrgan_model.py - UNet Discriminator
realesrgan/archs/discriminator_arch.py
Other components
- User
- Input Media
- Restored Media
- FFmpeg
- Training Data
Connections 25
- User → Image Inference: runs
- User → Video Inference: runs
- Image Inference → Input Media: reads
- Image Inference → RealESRGAN Upsampler: enhances
- Image Inference → Face Enhancer: optionally uses
- Image Inference → Restored Media: writes
- RealESRGAN Upsampler → Model Weights: loads
- RealESRGAN Upsampler → SRVGG Network: runs
- RealESRGAN Upsampler → BasicSR Networks: runs
- RealESRGAN Upsampler → Tile Processing: tiles
- Face Enhancer → RealESRGAN Upsampler: upsamples background
- Video Inference → Video Reader: starts
- Video Reader → Input Media: reads
- Video Reader → FFmpeg: decodes
- Video Inference → RealESRGAN Upsampler: restores frames
- Video Inference → Video Writer: starts
- Video Writer → FFmpeg: encodes
- Video Writer → Restored Media: writes
- Training Entry → Synthetic Dataset: configures
- Training Entry → Paired Dataset: configures
- Synthetic Dataset → Training Data: reads
- Paired Dataset → Training Data: reads
- Synthetic Dataset → Degradation Model: provides data
- Paired Dataset → Degradation Model: provides data
- Degradation Model → UNet Discriminator: optimizes