conda env create --file environment.yml

Alternatively, you may refer to StyleSDF environment setup.

Download the pre-trained models from Google drive link, create folder ./training_record/author_release/checkpoints and place the model under it. Download evaluate latents from Google drive link, place it under ./data.

python main.py  \
--jobname author_release --latents_eval latents_eval_50k.pth \
--exp_mode visualize_video --n_styles 11 --num_frames 250 \
--given_subject_list 1000-1010 --style_id 7 

NOTE: to manually change output directories before running NOTE: You can generate more data and save it as latents_eval.pth for visualization. The pre-trained models can be downloaded by running python download_models.py.

python generate_images_and_latents.py \
--style_field_option no_style_field --elastic_loss 0 --adaptive_style_mixing \
--output_root_dir data --identities 1000 

Run the following script to generate 10 styliized data corresponding to the real-space data generated above. NOTE: refer to DualStyle for more details.

bash generate_stylized_data.sh

The default style_data contains 10 styles and base style (real_space). Train the 1st 50 epochs without GAN loss for sake of speed.

python main.py  \
--jobname job_name \
--n_epoch 50

Continue to train 50 epochs with GAN loss

python main.py  \
--style_batch 1 \
--jobname job_name \
--n_epoch 50 \
--gan_loss 0.05 \
--continue_training 49

This code is built upon codebase of StyleSDF, and it also contains submodules including DualStyleGAN, VToonify, PerceptualSimilarity, and facexlib.

@inproceedings{zhang2023deformtoon3d,
 title = {DeformToon3D: Deformable 3D Toonification from Neural Radiance Fields},
 author = {Junzhe Zhang, Yushi Lan, Shuai Yang, Fangzhou Hong, Quan Wang, Chai Kiat Yeo, Ziwei Liu, Chen Change Loy},
 booktitle = {ICCV},
 year = {2023}}