How jantic/deoldify works
DeOldify is a deep-learning image and film colorization project. Users submit still images for artistic or stable colorization, or video that is split into frames, colorized frame-by-frame, and rebuilt. Training support includes image-to-image GAN/NoGAN workflows; fastai supplies the underlying data, vision, and learning framework. The README also describes pretrained-model usage and video stability through NoGAN training. The excerpts show the visualization entry point and video tools, but not all filter, generator, or training implementations, so uncertain call wiring is omitted.
Main entry points: visualize.py Image visualizer · dataset.py Training datasets · generators.py Generator models · basic_train.py Learner and training
Components 20
Colorization workflow
- Image visualizer
deoldify/visualize.py - Video colorizer
deoldify/visualize.py - Colorization filters
deoldify/filters.py - Result saving
deoldify/save.py
Models and training
- Training datasets
deoldify/dataset.py - Generator models
deoldify/generators.py - U-Net architecture
deoldify/unet.py - Custom layers
deoldify/layers.py - Augmentations
deoldify/augs.py - GAN critics
deoldify/critics.py - Training losses
deoldify/loss.py
Learning framework
- Learner and training
fastai/basic_train.py - Vision data pipeline
fastai/vision/data.py - Image transforms
fastai/vision/image.py - Training loop
fastai/train.py - Training callbacks
fastai/callback.py
Evaluation
- FID scoring
fid/fid_score.py - Inception features
fid/inception.py
Other components
- Image/video user
- FFmpeg
Connections 15
- Image/video user → Image visualizer: submits image
- Image/video user → Video colorizer: submits video
- Image visualizer → Colorization filters: applies filter
- Video colorizer → Image visualizer: colorizes frames
- Video colorizer → FFmpeg: extracts and rebuilds
- Image visualizer → Result saving: saves result
- Training datasets → Vision data pipeline: provides data
- Vision data pipeline → Image transforms: transforms images
- Learner and training → Training loop: runs training
- Learner and training → Training callbacks: dispatches callbacks
- Generator models → U-Net architecture: uses architecture
- Generator models → Custom layers: uses layers
- GAN critics → Custom layers: uses layers
- Training datasets → Augmentations: uses augmentation
- FID scoring → Inception features: extracts features