How davidcoach/retouchformer_aaai_24 works
RetouchFormer is a research implementation for high-quality face retouching: it predicts imperfection regions and synthesizes replacement content using clean-face priors and selective transformer attention. The documented workflow trains on paired FFHQR source/target images, with semi-supervised unpaired data, then evaluates or retouches images. The graph includes the training/data path and model components, plus documented capabilities. Some links reflect imports listed by the source index; unsampled inference wiring and several internal model interactions are left unasserted.
Main entry points: trainer.py Trainer · RetouchFormer.py Retouch generator · dataset.py Face datasets · train.py Training entry
Components 23
Training and evaluation
- Training entry
train.py - Trainer
core/trainer.py - Face datasets
core/dataset.py - Evaluation entry
eval.py - Image retouch entry
img_retouching.py
Retouching model
- Retouch generator
model/RetouchFormer.py - Prior and mask VQ
model/modules/soft_mask_generation.py - Window attention stages
model/network_vrt_pair_qkv.py - GPEN face prior
model/gpen_model.py - GAN discriminator
model/RetouchFormer.py - Spectral normalization
model/modules/spectral_norm.py
Training support
- Training losses
core/loss.py - Learning-rate scheduler
core/lr_scheduler.py - Quality metrics
core/metrics.py - Distributed setup
core/dist.py - IRSE face model
core/model_irse.py - Model helpers
core/helpers.py - Data utilities
core/utils.py - Window utilities
core/make_windows.py - Core spectral norm
core/spectral_norm.py
Model operations
- Fused activation
op/fused_act.py - Resampling operator
op/upfirdn2d.py
Other components
- Researcher
Connections 19
- Researcher → Training entry: starts training
README.md - Training entry → Trainer: starts
train.py - Trainer → Face datasets: loads datasets
core/trainer.py - Trainer → Training losses: uses losses
core/trainer.py - Trainer → Learning-rate scheduler: uses schedule
core/trainer.py - Evaluation entry → Face datasets: imports
eval.py - Image retouch entry → Face datasets: imports
img_retouching.py - Retouch generator → Prior and mask VQ: imports prior model
model/RetouchFormer.py - Retouch generator → Window attention stages: imports stages
model/RetouchFormer.py - Retouch generator → GPEN face prior: imports encoder/decoder
model/RetouchFormer.py - Retouch generator → Spectral normalization: imports normalization
model/RetouchFormer.py - Trainer → Retouch generator: trains retouch model
- Trainer → GAN discriminator: trains adversary
- Trainer → Quality metrics: reports quality
- Face datasets → Data utilities: imports utilities
core/dataset.py - Quality metrics → Data utilities: imports utilities
core/metrics.py - Training losses → IRSE face model: imports face model
core/loss.py - IRSE face model → Model helpers: imports helpers
core/model_irse.py - GPEN face prior → Resampling operator: imports resampling
model/gpen_model.py