This repository provides the official PyTorch implementation for the following paper:
Pastiche Master: Exemplar-Based High-Resolution Portrait Style Transfer
Shuai Yang, Liming Jiang, Ziwei Liu and Chen Change Loy
In CVPR 2022.
Project Page | Paper | Supplementary Video
Abstract: Recent studies on StyleGAN show high performance on artistic portrait generation by transfer learning with limited data. In this paper, we explore more challenging exemplar-based high-resolution portrait style transfer by introducing a novel DualStyleGAN with flexible control of dual styles of the original face domain and the extended artistic portrait domain. Different from StyleGAN, DualStyleGAN provides a natural way of style transfer by characterizing the content and style of a portrait with an intrinsic style path and a new extrinsic style path, respectively. The delicately designed extrinsic style path enables our model to modulate both the color and complex structural styles hierarchically to precisely pastiche the style example. Furthermore, a novel progressive fine-tuning scheme is introduced to smoothly transform the generative space of the model to the target domain, even with the above modifications on the network architecture. Experiments demonstrate the superiority of DualStyleGAN over state-of-the-art methods in high-quality portrait style transfer and flexible style control.
Features:
High-Resolution (1024) | Training Data-Efficient (~200 Images) | Exemplar-Based Color and Structure Transfer
- [02/2023] Add
--wplus
in style_transfer.py to use original w+ pSp encoder rather than z+. - [09/2022] Pre-trained models in three new styles (feat. StableDiffusion) are released.
- [07/2022] Source code license is updated.
- [03/2022] Paper and supplementary video are released.
- [03/2022] Web demo is created.
- [03/2022] Code is released.
- [03/2022] This website is created.
Integrated into Huggingface Spaces 🤗 using Gradio. Try out the Web Demo: or
Clone this repo:
git clone https://github.com/williamyang1991/DualStyleGAN.git
cd DualStyleGAN
Dependencies:
All dependencies for defining the environment are provided in environment/dualstylegan_env.yaml
.
We recommend running this repository using Anaconda:
conda env create -f ./environment/dualstylegan_env.yaml
We use CUDA 10.1 so it will install PyTorch 1.7.1 (corresponding to Line 22, Line 25, Line 26 of dualstylegan_env.yaml
). Please install PyTorch that matches your own CUDA version following https://pytorch.org/.
Cartoon, Caricature and Anime datasets can be downloaded from their official pages. We also provide the script to build new datasets.
Dataset | Description |
---|---|
Cartoon | 317 cartoon face images from Toonify. |
Caricature | 199 images from WebCaricature. Please refer to dataset preparation for more details. |
Anime | 174 images from Danbooru Portraits. Please refer to dataset preparation for more details. |
Fantasy | 137 fantasy face images generated by StableDiffusion. |
Illustration | 156 illustration face images generated by StableDiffusion. |
Impasto | 120 impasto face images generated by StableDiffusion. |
Other styles | Please refer to dataset preparation for the way of building new datasets. |
To help users get started, we provide a Jupyter notebook found in ./notebooks/inference_playground.ipynb
that allows one to visualize the performance of DualStyleGAN.
The notebook will download the necessary pretrained models and run inference on the images found in ./data/
.
If no GPU is available, you may refer to Inference on CPU, and set device = 'cpu'
in the notebook.
Pretrained models can be downloaded from Google Drive or Baidu Cloud (access code: cvpr):
Model | Description |
---|---|
encoder | Pixel2style2pixel encoder that embeds FFHQ images into StyleGAN2 Z+ latent code |
encoder_wplus | Original Pixel2style2pixel encoder that embeds FFHQ images into StyleGAN2 W+ latent code |
cartoon | DualStyleGAN and sampling models trained on Cartoon dataset, 317 (refined) extrinsic style codes |
caricature | DualStyleGAN and sampling models trained on Caricature dataset, 199 (refined) extrinsic style codes |
anime | DualStyleGAN and sampling models trained on Anime dataset, 174 (refined) extrinsic style codes |
arcane | DualStyleGAN and sampling models trained on Arcane dataset, 100 extrinsic style codes |
comic | DualStyleGAN and sampling models trained on Comic dataset, 101 extrinsic style codes |
pixar | DualStyleGAN and sampling models trained on Pixar dataset, 122 extrinsic style codes |
slamdunk | DualStyleGAN and sampling models trained on Slamdunk dataset, 120 extrinsic style codes |
fantasy | DualStyleGAN models trained on Fantasy dataset, 137 extrinsic style codes |
illustration | DualStyleGAN models trained on Illustration dataset, 156 extrinsic style codes |
impasto | DualStyleGAN models trained on Impasto dataset, 120 extrinsic style codes |
The saved checkpoints are under the following folder structure:
checkpoint
|--encoder.pt % Pixel2style2pixel model
|--encoder_wplus.pt % Pixel2style2pixel model (optional)
|--cartoon
|--generator.pt % DualStyleGAN model
|--sampler.pt % The extrinsic style code sampling model
|--exstyle_code.npy % extrinsic style codes of Cartoon dataset
|--refined_exstyle_code.npy % refined extrinsic style codes of Cartoon dataset
|--caricature
% the same files as in Cartoon
...
Transfer the style of a default Cartoon image onto a default face:
python style_transfer.py
The result cartoon_transfer_53_081680.jpg
is saved in the folder .\output\
,
where 53
is the id of the style image in the Cartoon dataset, 081680
is the name of the content face image.
An corresponding overview image cartoon_transfer_53_081680_overview.jpg
is additionally saved to illustrate the input content image, the encoded content image, the style image (* the style image will be shown only if it is in your folder), and the result:
Specify the style image with --style
and --style_id
(find the mapping between id and filename here, find the visual mapping between id and the style image here). Specify the filename of the saved images with --name
. Specify the weight to adjust the degree of style with --weight
.
The following script generates the style transfer results in the teaser of the paper.
python style_transfer.py
python style_transfer.py --style cartoon --style_id 10
python style_transfer.py --style caricature --name caricature_transfer --style_id 0 --weight 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
python style_transfer.py --style caricature --name caricature_transfer --style_id 187 --weight 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
python style_transfer.py --style anime --name anime_transfer --style_id 17 --weight 0 0 0 0 0.75 0.75 0.75 1 1 1 1 1 1 1 1 1 1 1
python style_transfer.py --style anime --name anime_transfer --style_id 48 --weight 0 0 0 0 0.75 0.75 0.75 1 1 1 1 1 1 1 1 1 1 1
Specify the content image with --content
. If the content image is not well aligned with FFHQ, use --align_face
. For preserving the color style of the content image, use --preserve_color
or set the last 11 elements of --weight
to all zeros.
python style_transfer.py --content ./data/content/unsplash-rDEOVtE7vOs.jpg --align_face --preserve_color \
--style arcane --name arcane_transfer --style_id 13 \
--weight 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 1 1 1 1 1 1 1
Specify --wplus
to use the original pSp encoder to extract the W+ intrinsic style code, which may better preserve the face features of the content image.
Remarks: Our trained pSp encoder on Z+/W+ space cannot perfectly encode the content image. If the style transfer result more consistent with the content image is desired, one may use latent optimization to better fit the content image or using other StyleGAN encoders (as discussed in williamyang1991#11 and williamyang1991#29).
More options can be found via python style_transfer.py -h
.
Generate random Cartoon face images (Results are saved in the ./output/
folder):
python generate.py
Specify the style type with --style
and the filename of the saved images with --name
:
python generate.py --style arcane --name arcane_generate
Specify the weight to adjust the degree of style with --weight
.
Keep the intrinsic style code, extrinsic color code or extrinsic structure code fixed using --fix_content
, --fix_color
and --fix_structure
, respectively.
python generate.py --style caricature --name caricature_generate --weight 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 --fix_content
More options can be found via python generate.py -h
.
Download the supporting models to the ./checkpoint/
folder:
Model | Description |
---|---|
stylegan2-ffhq-config-f.pt | StyleGAN model trained on FFHQ taken from rosinality. |
model_ir_se50.pth | Pretrained IR-SE50 model taken from TreB1eN for ID loss. |
Step 1: Prepare data. Prepare the dataset in ./data/DATASET_NAME/images/train/
. First create lmdb datasets:
python ./model/stylegan/prepare_data.py --out LMDB_PATH --n_worker N_WORKER --size SIZE1,SIZE2,SIZE3,... DATASET_PATH
For example, download 317 Cartoon images into ./data/cartoon/images/train/
and run
python ./model/stylegan/prepare_data.py --out ./data/cartoon/lmdb/ --n_worker 4 --size 1024 ./data/cartoon/images/
Step 2: Fine-tune StyleGAN. Fine-tune StyleGAN in distributed settings:
python -m torch.distributed.launch --nproc_per_node=N_GPU --master_port=PORT finetune_stylegan.py --batch BATCH_SIZE \
--ckpt FFHQ_MODEL_PATH --iter ITERATIONS --style DATASET_NAME --augment LMDB_PATH
Take the cartoon dataset for example, run (batch size of 8*4=32 is recommended)
python -m torch.distributed.launch --nproc_per_node=8 --master_port=8765 finetune_stylegan.py --iter 600 --batch 4 --ckpt ./checkpoint/stylegan2-ffhq-config-f.pt --style cartoon --augment ./data/cartoon/lmdb/
The fine-tuned model can be found in ./checkpoint/cartoon/finetune-000600.pt
. Intermediate results are saved in ./log/cartoon/
.
Step 3: Destylize artistic portraits.
python destylize.py --model_name FINETUNED_MODEL_NAME --batch BATCH_SIZE --iter ITERATIONS DATASET_NAME
Take the cartoon dataset for example, run:
python destylize.py --model_name finetune-000600.pt --batch 1 --iter 300 cartoon
The intrinsic and extrinsic style codes are saved in ./checkpoint/cartoon/instyle_code.npy
and ./checkpoint/cartoon/exstyle_code.npy
, respectively. Intermediate results are saved in ./log/cartoon/destylization/
.
To speed up destylization, set --batch
to large value like 16.
For styles severely different from real faces, set --truncation
to small value like 0.5 to make the results more photo-realistic (it enables DualStyleGAN to learn larger structrue deformations).
Stage 1 & 2: Pretrain DualStyleGAN on FFHQ. We provide our pretrained model generator-pretrain.pt at Google Drive or Baidu Cloud (access code: cvpr). This model is obtained by:
python -m torch.distributed.launch --nproc_per_node=1 --master_port=8765 pretrain_dualstylegan.py --iter 3000 --batch 4 ./data/ffhq/lmdb/
where ./data/ffhq/lmdb/
contains the lmdb data created from the FFHQ dataset via ./model/stylegan/prepare_data.py
.
Stage 3: Fine-Tune DualStyleGAN on Target Domain. Fine-tune DualStyleGAN in distributed settings:
python -m torch.distributed.launch --nproc_per_node=N_GPU --master_port=PORT finetune_dualstylegan.py --iter ITERATIONS \
--batch BATCH_SIZE --ckpt PRETRAINED_MODEL_PATH --augment DATASET_NAME
The loss term weights can be specified by --style_loss
(λFM), --CX_loss
(λCX), --perc_loss
(λperc), --id_loss
(λID) and --L2_reg_loss
(λreg). λID and λreg are suggested to be tuned for each style dataset to achieve ideal performance. More options can be found via python finetune_dualstylegan.py -h
.
Take the Cartoon dataset as an example, run (multi-GPU enables a large batch size of 8*4=32 for better performance):
python -m torch.distributed.launch --nproc_per_node=8 --master_port=8765 finetune_dualstylegan.py --iter 1500 --batch 4 --ckpt ./checkpoint/generator-pretrain.pt --style_loss 0.25 --CX_loss 0.25 --perc_loss 1 --id_loss 1 --L2_reg_loss 0.015 --augment cartoon
The fine-tuned models can be found in ./checkpoint/cartoon/generator-ITER.pt
where ITER = 001000, 001100, ..., 001500. Intermediate results are saved in ./log/cartoon/
. Large ITER has strong cartoon styles but at the cost of artifacts, and users may select the most balanced one from 1000-1500. We use 1400 for our paper experiments.
Refine extrinsic style code. Refine the color and structure styles to better fit the example style images.
python refine_exstyle.py --lr_color COLOR_LEARNING_RATE --lr_structure STRUCTURE_LEARNING_RATE DATASET_NAME
By default, the code will load instyle_code.npy
, exstyle_code.npy
, and generator.pt
in ./checkpoint/DATASET_NAME/
. Use --instyle_path
, --exstyle_path
, --ckpt
to specify other saved style codes or models. Take the Cartoon dataset as an example, run:
python refine_exstyle.py --lr_color 0.1 --lr_structure 0.005 --ckpt ./checkpoint/cartoon/generator-001400.pt cartoon
The refined extrinsic style codes are saved in ./checkpoint/DATASET_NAME/refined_exstyle_code.npy
. lr_color
and lr_structure
are suggested to be tuned to better fit the example styles.
Training sampling network. Train a sampling network to map unit Gaussian noises to the distribution of extrinsic style codes:
python train_sampler.py DATASET_NAME
By default, the code will load refined_exstyle_code.npy
or exstyle_code.npy
in ./checkpoint/DATASET_NAME/
. Use --exstyle_path
to specify other saved extrinsic style codes. The saved model can be found in ./checkpoint/DATASET_NAME/sampler.pt
.
cartoon.mp4
caricature.mp4
anime.mp4
We use StableDiffusion to generate face images of the specified style of famous artists. Trained with these images, DualStyleGAN is able to pastiche these famous artists and generates appealing results.
If you find this work useful for your research, please consider citing our paper:
@inproceedings{yang2022Pastiche,
title={Pastiche Master: Exemplar-Based High-Resolution Portrait Style Transfer},
author={Yang, Shuai and Jiang, Liming and Liu, Ziwei and Loy, Chen Change},
booktitle={CVPR},
year={2022}
}
The code is mainly developed based on stylegan2-pytorch and pixel2style2pixel.