🧩 TokenCompose: Text-to-Image Diffusion with Token-level Supervision

Zirui Wang^{1, 3} · Zhizhou Sha^{2, 3} · Zheng Ding³ · Yilin Wang^{2, 3} · Zhuowen Tu³

¹Princeton University · ²Tsinghua University · ³University of California, San Diego

CVPR 2024

Project done while Zirui Wang, Zhizhou Sha and Yilin Wang interned at UC San Diego.

Project Page | arXiv | X (Twitter)

Updates

If you use our method and/or model for your research project, we are happy to provide cross-reference here in the updates. :)

[04/04/2024] 🔥 Our training methodology is incorporated into CoMat which shows enhanced text-to-image attribute assignments.
[02/26/2024] 🔥 TokenCompose is accepted to CVPR 2024!
[02/20/2024] 🔥 TokenCompose is used as a base model from the RealCompo paper for enhanced compositionality.

https://github.com/mlpc-ucsd/TokenCompose/assets/59942464/93feea16-4eac-49c3-b286-ee390a325b17

A Stable Diffusion model finetuned with token-level consistency terms for enhanced multi-category instance composition and photorealism.

🆕 Models

Our finetuned models do not contain any extra modules and can be directly used in a standard diffusion model library (e.g., HuggingFace's Diffusers) by replacing the pretrained U-Net with our finetuned U-Net in a plug-and-play manner. We provide a demo jupyter notebook which uses our model checkpoint to generate images.

You can also use the following code to download our checkpoints and generate images:

import torch
from diffusers import StableDiffusionPipeline

model_id = "mlpc-lab/TokenCompose_SD14_A"
device = "cuda"

pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float32)
pipe = pipe.to(device)

prompt = "A cat and a wine glass"
image = pipe(prompt).images[0]  
    
image.save("cat_and_wine_glass.png")

📊 MultiGen

See MultiGen for details.

💻 Environment Setup

For those who want to use our codebase to train your own diffusion models with token-level objectives, follow the below instructions:

conda create -n TokenCompose python=3.8.5
conda activate TokenCompose
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia
pip install -r requirements.txt

We have verified the environment setup using this specific package versions, but we expect that it will also work for newer versions too!

🛠️ Dataset Setup

If you want to use your own data, please refer to preprocess_data for details.

If you want to use our training data as examples or for research purposes, please follow the below instructions:

1. Setup the COCO Image Data

cd train/data
# download COCO train2017
wget http://images.cocodataset.org/zips/train2017.zip
unzip train2017.zip
rm train2017.zip
bash coco_data_setup.sh

After this step, you should have the following structure under the train/data directory:

train/data/
    coco_gsam_img/
        train/
            000000000142.jpg
            000000000370.jpg
            ...

2. Setup Token-wise Grounded Segmentation Maps

Download COCO segmentation data from Google Drive and put it under train/data directory.

After this step, you should have the following structure under the train/data directory:

train/data/
    coco_gsam_img/
        train/
            000000000142.jpg
            000000000370.jpg
            ...
    coco_gsam_seg.tar

Then, run the following command to unzip the segmentation data:

cd train/data
tar -xvf coco_gsam_seg.tar
rm coco_gsam_seg.tar

After the setup, you should have the following structure under the train/data directory:

train/data/
    coco_gsam_img/
        train/
            000000000142.jpg
            000000000370.jpg
            ...
    coco_gsam_seg/
        000000000142/
            mask_000000000142_bananas.png
            mask_000000000142_bread.png
            ...
        000000000370/
            mask_000000000370_bananas.png
            mask_000000000370_bread.png
            ...
        ...

📈 Training

We use wandb to log some curves and visualizations. Login to wandb before running the scripts.

wandb login

Then, to run TokenCompose, use the following command:

cd train
bash train.sh

The results will be saved under train/results directory.

🏷️ License

This repository is released under the Apache 2.0 license.

🙏 Acknowledgement

Our code is built upon diffusers, prompt-to-prompt, VISOR, Grounded-Segment-Anything, and CLIP. We thank all these authors for their nicely open sourced code and their great contributions to the community.

📝 Citation

If you find our work useful, please consider citing:

@InProceedings{Wang2024TokenCompose,
    author    = {Wang, Zirui and Sha, Zhizhou and Ding, Zheng and Wang, Yilin and Tu, Zhuowen},
    title     = {TokenCompose: Text-to-Image Diffusion with Token-level Supervision},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {8553-8564}
}