hjinnkim / StyleKD

[ECCV2022] Mind the Gap in Distilling StyleGANs

StyleKD

This repo is the implementation of paper Mind the Gap in Distilling StyleGANs (ECCV2022).

Running

This is unstable version. The code is still in testing.

CUDA_VISIBLE_DEVICES=4,5,6,7 python3 -m torch.distributed.launch --nproc_per_node=4 --master_port=8001 distributed_train.py --name batch16_run1 \
    --load 0 --load_style 1 --g_step 0 \
    --kd_l1_lambda 3 --kd_lpips_lambda 3 --kd_simi_lambda 30 --kd_style_lambda 0 \
    --lr_mlp 0.01 --fix_w 0 --batch 4 --worker 4 \
    --simi_loss kl --single_view 0 --offset_mode main --main_direction split --offset_weight 5.0

Results

Quantitative Results

Qualitative Results

Face

Church

Face Editing

Citation

If you find this repo useful for your research, please consider citing the paper

@misc{xu2022stylekd,
  url = {https://arxiv.org/abs/2208.08840},
  author = {Xu, Guodong and Hou, Yuenan and Liu, Ziwei and Loy, Chen Change},
  keywords = {Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {Mind the Gap in Distilling StyleGANs},
  publisher = {arXiv},
  year = {2022}
}