How ewrfcas/zits-plusplus works
ZITS++ is an image-inpainting system: users provide images and masks, the runtime prepares inputs and structural guidance, then a learned generator reconstructs masked regions for output. The README documents 256/512 inference, optional object removal, pretrained models, and training-related code. Sampled dataset, trainer, generator, and line-detector code supports the main stages; the inference entry point and some model wiring are absent from the supplied tree or excerpts, so uncertain connections are omitted.
Main entry points: pl_trainers.py Inpainting trainer · dataloader.py Test dataset · dataset.py Training datasets · detector.py LSM-HAWP detector
Components 15
Input preparation
- Test dataset
dataset/dataloader.py - Training datasets
base/dataset.py
Structural priors
- LSM-HAWP detector
trainers/lsm_hawp/detector.py - ADE20K segmentation
networks/ade20k/base.py - Structure upsampler
networks/tsr.py - Prior networks
networks/pcp.py
Inpainting models
- MAT generator
networks/mat.py - Generator components
networks/generators.py - Discriminator
networks/discriminators.py - Training losses
networks/losses.py
Runtime and training
- Inpainting trainer
trainers/pl_trainers.py - Inpainting metrics
inpainting_metric.py - Training logger
logger/logger.py
Other components
- Image and mask user
- Inpainted image
Connections 14
- Image and mask user → Test dataset: provides images/masks
- Test dataset → Inpainting trainer: supplies batches
- Training datasets → Inpainting trainer: supplies batches
- Inpainting trainer → LSM-HAWP detector: uses line prior
- Inpainting trainer → Structure upsampler: uses structure prior
- Inpainting trainer → Prior networks: uses priors
- Inpainting trainer → MAT generator: runs generator
- Inpainting trainer → Generator components: runs generator
- Inpainting trainer → Discriminator: scores images
- Inpainting trainer → Training losses: computes losses
- Inpainting trainer → Inpainted image: produces image
- Inpainting trainer → Training logger: logs progress
- Inpainted image → Inpainting metrics: evaluated by
- Inpainting trainer → ADE20K segmentation: optional prior