arXiv paper

We propose the two-stream convolution and the local autoregressive mask to maintain the image consistence without information leakage.

• Release the training codes.
• Codes for pose editing.

These codes are only tested in pytorch==1.3.1.

1. Download the pretrained sketch-vqgan, image-vqgan, and transformer.
2. For masked training, we provide irregular and segmentation masks (download) with different masking rates. And you should define the mask file list before the training (flist_example.txt).

The TS-VQGAN is designed a little different from VQGAN, which can trained with less GPU memory with the same performance.

1. Remove attentions.

2. Using InstanceNorm instead of GroupNorm.

3. Progressive channel widths.

4. Fewer parameters for the sketch vqgan.

1. Generating XDoG sketches.

python generate_xdog_sketch.py --input_path <img fold> --output_path <output fold>

2. Train sketch VQGAN.

python train_sketch_vqgan.py --path <model path> --config_path configs/vqgan_ffhq.yml --max_iters 150000 --learning_rate 1e-4 --gpu 0

3. Train image TS-VQGAN.

python train_image_tsvqgan.py --path <model path> --config_path configs/vqgan_ffhq.yml --max_iters 150000 --learning_rate 2e-4 --gpu 0
python train_image_tsvqgan.py --path <model path> --config_path configs/vqgan_ffhq.yml --max_iters 300000 --learning_rate 4e-5 --gpu 0 --finetune

4. Train the iLAT.

python train_transformer.py --path <model path> --config_path configs/transformer_ffhq.yml \
                            --sketch_model_path <sketch-VQGAN path> \
                            --image_model_path <image-VQGAN path> \
                            --max_iters 300000 --learning_rate 5e-5 --gpu 3

See face_editing_demo.ipynb.