BiGR: Harnessing Binary Latent Codes for Image Generation and Improved Visual Representation Capabilities

This is the official PyTorch code for the paper:

BiGR: Harnessing Binary Latent Codes for Image Generation and Improved Visual Representation Capabilities
Shaozhe Hao¹, Xuantong Liu², Xianbiao Qi³*, Shihao Zhao¹, Bojia Zi⁴, Rong Xiao³, Kai Han¹†, Kwan-Yee K. Wong¹†
¹The University of Hong Kong   ²Hong Kong University of Science and Technology
³Intellifusion   ⁴The Chinese University of Hong Kong
(*: Project lead; †: Corresponding authors)

[Project page] [arXiv] [Colab]

TL;DR: We introduce BiGR, a novel conditional image generation model using compact binary latent codes for generative training, focusing on enhancing both generation and representation capabilities.

📢 News

🌟 We are training BiGR with REPA, a representation alignment regularization that enhances both generation and representation performance in DiT/SiT.

⚙️ Setup

You can simply install the environment with the file environment.yml by:

conda env create -f environment.yml
conda activate BiGR

🔗 Download

Please first download the pretrained weights for tokenizers and BiGR models to run our tests.

Binary Autoencoder

We train Binary Autoencoder (B-AE) by adapting the official code of Binary Latent Diffusion. We provide pretrained weights for different configurations.

256x256 resolution

B-AE	Size	Checkpoint
d24	332M	download
d32	332M	download

512x512 resolution

B-AE	Size	Checkpoint
d32-512	315M	download

BiGR models ✨

We provide pretrained weights for BiGR models in various sizes.

256x256 resolution

Model	B-AE	Size	Checkpoint
BiGR-L-d24	d24	1.35G	download
BiGR-XL-d24	d24	3.20G	download
BiGR-XXL-d24	d24	5.92G	download
BiGR-XXL-d32	d32	5.92G	download

512x512 resolution

Model	B-AE	Size	Checkpoint
BiGR-L-d32-res512	d32-res512	1.49G	download

🚀 Image generation

We provide the sample script for 256x256 image generation in script/sample.sh.

bash script/sample.sh

Please specify the code dimension $CODE, your B-AE checkpoint path $CKPT_BAE, and your BiGR checkpoint path $CKPT_BIGR.

You may also want to try different settings of the CFG scale $CFG, the number of sample iterations $ITER, and the gumbel temperature $GUMBEL. We recommend using small gumbel temperature for better visual quality (e.g., GUMBEL=0). You can increase gumbel temperature to enhance generation diversity.

You can generate 512x512 images using script/sample_512.sh. Note that you need to specify the corresponding 512x512 tokenizers and models.

bash script/sample_512.sh

💡 Zero-shot applications

BiGR supports various zero-shot generalized applications, without the need for task-specific structural changes or parameter fine-tuning.

You can easily download testing images and run our scripts to get started. Feel free to play with your own images.

Inpainting & Outpainting

bash script/app_inpaint.sh

bash script/app_outpaint.sh

You need to save the source image and the mask in the same folder, with the image as a *.JPEG file and the mask as a *.png file. You can then specify the source image path $IMG.

You can customize masks using this gradio demo.

Class-conditional editting

bash script/app_edit.sh

In addition to the source image path $IMG, you also need to give a class index $CLS for editing.

Class interpolation

bash script/app_interpolate.sh

You need to specify two class indices $CLS1 and $CLS2.

Image enrichment

bash script/app_enrich.sh

You need to specify the source image path $IMG.

💻 Train

You can train BiGR yourself by running:

bash script/train.sh

You need to specify the ImageNet-1K dataset path --data-path.

We train L/XL-sized models using 8 A800 GPUs and XXL-sized models using 32 A800 GPUs on 4 nodes.

💐 Acknowledgement

This project builds on Diffusion Transformer, Binary Latent Diffusion, and LlamaGen. We thank these great works!

📖 Citation

If you use this code in your research, please consider citing our paper:

@misc{hao2024bigr,
    title={Bi{GR}: Harnessing Binary Latent Codes for Image Generation and Improved Visual Representation Capabilities}, 
    author={Shaozhe Hao and Xuantong Liu and Xianbiao Qi and Shihao Zhao and Bojia Zi and Rong Xiao and Kai Han and Kwan-Yee~K. Wong},
    year={2024},
}