chenwu98 / unified-generative-zoo

[ICCV 2023] https://arxiv.org/abs/2210.05559

A Unified Interface for Guiding Generative Models (2D/3D GANs, Diffusion Models, and Their Variants)

Official PyTorch implementation of (Section 4.3 of) our paper
Unifying Diffusion Models' Latent Space, with Applications to CycleDiffusion and Guidance
Chen Henry Wu, Fernando De la Torre
Carnegie Mellon University
Preprint, Oct 2022

[Paper link]

Updates

[Oct 13, 2022] Code released.

Notes

1. Sections 4.1 and 4.2 of this paper is open-sourced at CycleDiffusion.
2. The code is based on Generative Visual Prompt.
3. Feel free to email me if you think I should cite your work!

Overview

GANs, VAEs, and normalizing flows are usually characterized as deterministic mappings from isometric Gaussian latent codes to images. We show that it is possible to unify various diffusion models into this formulation. This allows us to guide (or condition, control) various generative models in a unified, plug-and-play manner by leveraging latent-space energy-based models (EBMs). This repository provides a unified interface for guiding various generative models with CLIP, classifiers, and face IDs.

Models studied in this paper (some of them are not included here; please check CycleDiffusion):

An illustration of generative models as deterministic mappings from isometric Gaussian latent codes to images.

Interestingly, we find that different models represent subpopulations and individuals in different ways, although most of them are trained on the same data.

Contents

• A Unified Interface for Guiding Generative Models (2D/3D GANs, Diffusion Models, and Their Variants)
  • Updates
  • TODOs
  • Notes
  • Overview
  • Contents
  • Dependencies
  • Pre-trained checkpoints
    • Pre-trained generative models
    • Off-the-shelf models for guidance
  • Usage
    • Overview
    • CLIP guidance for sampling sub-populations
    • Classifier guidance for sampling sub-populations
    • ID guidance for sampling individuals
  • Citation
  • License
  • Contact

Dependencies

1. Create environment by running

conda env create -f environment.yml
conda activate generative_prompt
pip install git+https://github.com/openai/CLIP.git

2. Install torch and torchvision based on your CUDA version.
3. Install PyTorch 3D. Installing this library can be painful, but you can skip it if you are not using 3D GANs.
4. Install taming-transformers by running

cd ../
git clone [email protected]:CompVis/taming-transformers.git
cd taming-transformers/
pip install -e .
cd ../

5. Set up wandb for logging (registration is required). You should modify the setup_wandb function in main.py to accomodate your wandb credentials. You may want to run something like

wandb login

Pre-trained checkpoints

Pre-trained generative models

We provide a unified interface for various pre-trained generative models. Checkpoints for generative models used in this paper are provided below.

1. StyleGAN2

cd ckpts/
wget https://www.dropbox.com/s/iy0dkqnkx7uh2aq/ffhq.pt
wget https://www.dropbox.com/s/lmjdijm8cfmu8h1/metfaces.pt
wget https://www.dropbox.com/s/z1vts069w683py5/afhqcat.pt
wget https://www.dropbox.com/s/a0hvdun57nvafab/stylegan2-church-config-f.pt
wget https://www.dropbox.com/s/x1d19u8zd6yegx9/stylegan2-car-config-f.pt
wget https://www.dropbox.com/s/hli2x42ekdaz2br/landscape.pt

2. StyleNeRF

cd ckpts/
wget https://www.dropbox.com/s/dtqsroh95uquwoc/StyleNeRF_ffhq_256.pkl
wget https://www.dropbox.com/s/klbuhqfv74q7e35/StyleNeRF_ffhq_512.pkl
wget https://www.dropbox.com/s/n80cr7isveh5yfu/StyleNeRF_ffhq_1024.pkl

3. Extended Analytic DPM

cd ckpts/
mkdir extended_adpm
cd extended_adpm/
wget https://www.dropbox.com/s/r8210seh6ekhogf/celeba64_ema_eps_epsc_pretrained_190000.ckpt.pth
wget https://www.dropbox.com/s/6o5etzhgbihr0yh/celeba64_ema_eps_eps2_pretrained_340000.ckpt.pth
wget https://www.dropbox.com/s/o0jw5ezai1e1z3v/celeba64_ema_eps.ckpt.pth
wget https://www.dropbox.com/s/0axtykkvyz49hrw/celeba64_ema_eps.ms_eps.pth

4. StyleGAN-XL

# StyleGAN-XL will be downloaded automatically. 

5. StyleSwin

cd ckpts/
wget https://www.dropbox.com/s/f0nlvu6fh3bbpmd/StyleSwin_FFHQ_1024.pt
wget https://www.dropbox.com/s/c2812gumbyxj751/StyleSwin_FFHQ_256.pt

6. StyleSDF

cd ckpts/
wget https://www.dropbox.com/s/epet782zdu0hazx/stylesdf_ffhq_vol_renderer.pt
wget https://www.dropbox.com/s/p0ptofh7sku2o8j/stylesdf_ffhq1024x1024.pt
wget https://www.dropbox.com/s/rq756clx14a9kgd/stylesdf_afhq_vol_renderer.pt
wget https://www.dropbox.com/s/hu5wgr40vyptzx6/stylesdf_afhq512x512.pt
wget https://www.dropbox.com/s/8rsaxzmey64jugo/stylesdf_sphere_init.pt

7. Diffusion Autoencoder

cd ckpts/
wget https://www.dropbox.com/s/ej0jj8g7crvtb5e/diffae_ffhq256.ckpt
wget https://www.dropbox.com/s/w5y89y57r9nd1jt/diffae_ffhq256_latent.pkl
wget https://www.dropbox.com/s/rsbpxaswnfzsyl1/diffae_ffhq128.ckpt
wget https://www.dropbox.com/s/v1dvsj6oklpz652/diffae_ffhq128_latent.pkl

8. Latent Diffusion Model

cd ckpts/
wget https://www.dropbox.com/s/9lpdgs83l7tjk6c/ldm_models.zip
unzip ldm_models.zip

9. NVAE

cd ckpts/
wget https://www.dropbox.com/s/bwwtszb5g5alw30/nvae_ffhq_256.pt
wget https://www.dropbox.com/s/8dfryaandkmoxzz/nvae_celebahq_256.pt

10. EG3D

cd ckpts/
wget --content-disposition https://api.ngc.nvidia.com/v2/models/nvidia/research/eg3d/versions/1/zip -O eg3d.zip
unzip eg3d.zip

11. Denoising Diffusion GAN

cd ckpts/
wget https://www.dropbox.com/s/lsfkbln9u78rbhs/ddgan_celebahq256_netG_550.pth

12. GIRAFFE-HQ

cd ckpts/
wget https://www.dropbox.com/s/jj03hto6o9rnbha/giraffehd_ffhq_1024.pt

13. Diffusion-GAN

cd ckpts/
wget https://www.dropbox.com/s/25ryma8et4ohmjq/diffusion-stylegan2-ffhq.pkl

Off-the-shelf models for guidance

1. CLIP

# CLIP will be downloaded automatically

2. ArcFace IR-SE 50 model, provided by the Colab demo in this repo

cd ckpts/
wget https://www.dropbox.com/s/qg7co4azsv5sacm/model_ir_se50.pth

3. CelebA classifier, trained by this repo

cd ckpts/
wget https://www.dropbox.com/s/yzc8ydaa4ggj1zs/celeba.zip
unzip celeba.zip 

Usage

Overview

Each set notation {A,B,C} stands for several independent experiments. You should always replace {A,B,C} with one of A, B, and C. Model checkpoints and image samples will be saved under --output_dir.

CLIP guidance for sampling sub-populations

1. Generative model $\in$ {LDM-DDIM, DiffAE, Diffusion-GAN, StyleGAN-XL, StyleGAN2, StyleNeRF, StyleSDF, EG3D, GIRAFFE-HD, StyleSwin, NVAE}.
2. Dataset and resolution $\in$ {FFHQ1024, FFHQ512, FFHQ256, FFHQ128}.
3. Text description $\in$ {"a photo of a baby", "a photo of an old person", "a photo of a person with eyeglasses", "a photo of a person with eyeglasses and a yellow hat"}.
4. Guidance strength $\lambda_{\text{CLIP}}$ $\in$ {100, 300, 500, 700, 1000}. In the following command, _500 is omitted.
5. Note that not all combinations of Generative model $\times$ Dataset and resolution are available. Please check the paper and available configs for details.

export CUDA_VISIBLE_DEVICES=0
export RUN_NAME=clip_{a_baby,an_old_person,a_person_with_eyeglasses,a_person_with_eyeglasses_and_a_yellow_hat}_{ffhq1024,ffhq512,ffhq256,ffhq128}_{styleganxl,stylegan2,styleswin,stylenerf,latentdiff_5step,latentdiff_10step,diffae_3step_3step_latent_only,stylesdf,stylegan2_no_trunc,stylesdf_no_trunc,styleswin_no_trunc,styleganxl_no_trunc,stylenerf_no_trunc,nvae,eg3d,eg3d_no_trunc,giraffehd,diffae_3step_3step_both,diffusion_stylegan2,diffusion_stylegan2_no_trunc,diffae_10step_10step_both,}_langevin{,_100,_300,_700,_1000}
export SEED=42
nohup python -m torch.distributed.launch --nproc_per_node 1 --master_port 1410 main.py --seed $SEED --cfg experiments/$RUN_NAME.cfg --run_name $RUN_NAME$SEED --logging_strategy steps --logging_first_step true --logging_steps 4 --evaluation_strategy steps --eval_steps 50 --metric_for_best_model CLIPEnergy --greater_is_better false --save_strategy steps --save_steps 50 --save_total_limit 1 --load_best_model_at_end --gradient_accumulation_steps 4 --num_train_epochs 0 --adafactor false --learning_rate 1e-3 --do_eval --output_dir output/$RUN_NAME$SEED --overwrite_output_dir --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --eval_accumulation_steps 4 --ddp_find_unused_parameters true --verbose true > $RUN_NAME$SEED.log 2>&1 &

Classifier guidance for sampling sub-populations

# DDGAN CelebAHQ256 old
export CUDA_VISIBLE_DEVICES=0
export RUN_NAME=class_old_celebahq256_ddgan_langevin
export SEED=42
nohup python -m torch.distributed.launch --nproc_per_node 1 --master_port 1430 main.py --seed $SEED --cfg experiments/$RUN_NAME.cfg --run_name $RUN_NAME$SEED --logging_strategy steps --logging_first_step true --logging_steps 4 --evaluation_strategy steps --eval_steps 50 --metric_for_best_model ClassEnergy --greater_is_better false --save_strategy steps --save_steps 50 --save_total_limit 1 --load_best_model_at_end --gradient_accumulation_steps 4 --num_train_epochs 0 --adafactor false --learning_rate 1e-3 --do_eval --output_dir output/$RUN_NAME$SEED --overwrite_output_dir --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --eval_accumulation_steps 4 --ddp_find_unused_parameters true --verbose true > $RUN_NAME$SEED.log 2>&1 &

# DDGAN CelebAHQ256 eyeglassess
export CUDA_VISIBLE_DEVICES=0
export RUN_NAME=class_eyeglasses_celebahq256_ddgan_langevin
export SEED=42
nohup python -m torch.distributed.launch --nproc_per_node 1 --master_port 1430 main.py --seed $SEED --cfg experiments/$RUN_NAME.cfg --run_name $RUN_NAME$SEED --logging_strategy steps --logging_first_step true --logging_steps 4 --evaluation_strategy steps --eval_steps 50 --metric_for_best_model ClassEnergy --greater_is_better false --save_strategy steps --save_steps 50 --save_total_limit 1 --load_best_model_at_end --gradient_accumulation_steps 4 --num_train_epochs 0 --adafactor false --learning_rate 1e-3 --do_eval --output_dir output/$RUN_NAME$SEED --overwrite_output_dir --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --eval_accumulation_steps 4 --ddp_find_unused_parameters true --verbose true > $RUN_NAME$SEED.log 2>&1 &

# SN-DPM DDPM CelebA64 old
export CUDA_VISIBLE_DEVICES=0
export RUN_NAME=class_old_celeba64_sn_dpm_ddpm_langevin
export SEED=42
nohup python -m torch.distributed.launch --nproc_per_node 1 --master_port 1430 main.py --seed $SEED --cfg experiments/$RUN_NAME.cfg --run_name $RUN_NAME$SEED --logging_strategy steps --logging_first_step true --logging_steps 4 --evaluation_strategy steps --eval_steps 50 --metric_for_best_model ClassEnergy --greater_is_better false --save_strategy steps --save_steps 50 --save_total_limit 1 --load_best_model_at_end --gradient_accumulation_steps 4 --num_train_epochs 0 --adafactor false --learning_rate 1e-3 --do_eval --output_dir output/$RUN_NAME$SEED --overwrite_output_dir --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --eval_accumulation_steps 4 --ddp_find_unused_parameters true --verbose true > $RUN_NAME$SEED.log 2>&1 &

# SN-DPM DDPM CelebA64 eyeglasses
export CUDA_VISIBLE_DEVICES=0
export RUN_NAME=class_eyeglasses_celeba64_sn_dpm_ddpm_langevin
export SEED=42
nohup python -m torch.distributed.launch --nproc_per_node 1 --master_port 1430 main.py --seed $SEED --cfg experiments/$RUN_NAME.cfg --run_name $RUN_NAME$SEED --logging_strategy steps --logging_first_step true --logging_steps 4 --evaluation_strategy steps --eval_steps 50 --metric_for_best_model ClassEnergy --greater_is_better false --save_strategy steps --save_steps 50 --save_total_limit 1 --load_best_model_at_end --gradient_accumulation_steps 4 --num_train_epochs 0 --adafactor false --learning_rate 1e-3 --do_eval --output_dir output/$RUN_NAME$SEED --overwrite_output_dir --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --eval_accumulation_steps 4 --ddp_find_unused_parameters true --verbose true > $RUN_NAME$SEED.log 2>&1 &

ID guidance for sampling individuals

1. ID reference $\in$ {00001, 00002, 00015, 00018}.
2. Generative model $\in$ {LDM-DDIM, DiffAE, StyleGAN-XL, StyleGAN2, DDGAN, EG3D, GIRAFFE-HD}.
3. Dataset and resolution $\in$ {FFHQ1024, FFHQ512, FFHQ256, CelebAHQ256}.
4. Note that not all combinations of Generative model $\times$ Dataset and resolution are available. Please check the paper and available configs for details.

export CUDA_VISIBLE_DEVICES=0
export RUN_NAME=recon_id_{00001,00002,00015,00018}_{ffhq256,ffhq512,ffhq1024,celebahq256}_{latentdiff,diffae_10steps_10steps_both,giraffehd,stylegan2,styleganxl,ddgan,eg3d}_langevin
export SEED=42
nohup python -m torch.distributed.launch --nproc_per_node 1 --master_port 1420 main.py --seed $SEED --cfg experiments/$RUN_NAME.cfg --run_name $RUN_NAME$SEED --logging_strategy steps --logging_first_step true --logging_steps 4 --evaluation_strategy steps --eval_steps 50 --metric_for_best_model CLIPEnergy --greater_is_better false --save_strategy steps --save_steps 50 --save_total_limit 1 --load_best_model_at_end --gradient_accumulation_steps 8 --num_train_epochs 0 --adafactor false --learning_rate 1e-3 --do_eval --output_dir output/$RUN_NAME$SEED --overwrite_output_dir --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --eval_accumulation_steps 4 --ddp_find_unused_parameters true --verbose true > $RUN_NAME$SEED.log 2>&1 &

Citation

If you find this repository helpful, please cite as

@inproceedings{unifydiffusion2022,
  title={Unifying Diffusion Models' Latent Space, with Applications to {CycleDiffusion} and Guidance},
  author={Chen Henry Wu and Fernando De la Torre},
  booktitle={ArXiv},
  year={2022},
}

Do not forget to cite the original papers that proposed these models!

License

We use the X11 License. This license is identical to the MIT License, but with an extra sentence that prohibits using the copyright holders' names (Carnegie Mellon University in our case) for advertising or promotional purposes without written permission.

Contact

Issues are welcome if you have any question about the code. If you would like to discuss the method, please contact Chen Henry Wu.

unified-generative-zoo's Issues

Computation of Langevin dynamics

In the line100 in langevin_dynamics.py, the z is updated according to the Langevin dynamics:
z = (z - self.step_size / 2 * grad + noise).detach().
The variable `grad' in code is computed as $\bigtriangledown_z -E(G(z)|C)$.

Refer to Sec 3.4 【UNIFIED PLUG-AND-PLAY GUIDANCE FOR GENERATIVE MODELS】 in paper 【UNIFYING DIFFUSION MODELS’ LATENT SPACE, WITH APPLICATIONS TO CYCLEDIFFUSION AND GUIDANCE】, the update of z should involve the term $\bigtriangledown_z \log(p_z(z^k))$.

I am wondering why this term can be omitted in practice or which codes compute this term. Can you give me some hints? Thank you for your great work and your kindly reply.