Home

Awesome

[ECCV 2024] Powerful and Flexible: Personalized Text-to-Image Generation via Reinforcement Learning

<div align="center">

<a href='https://arxiv.org/abs/2407.06642v2'><img src='https://img.shields.io/badge/arXiv-2407.06642-b31b1b.svg'></a>  

</div>

🔆 Introduction

This repo contains the official code of our ECCV2024 paper: [Powerful and Flexible: Personalized Text-to-Image Generation via Reinforcement Learning]

The paper and code will be release soon in the next 1~2 weeks.

⚙️ Setup

Before running the script, make sure you install the library from source:

git clone https://github.com/huggingface/diffusers
cd diffusers
pip install .
pip install -r requirements.txt

💥 Training

Using 'Look Forward' reward

Take backpack_dog(backpack) as example. Put your pretrained model in path/to/pretrained_stable_diffusion, We use Stable-Diffusion-V1.4 in our paper.

Put your personalized collections in path/to/personalized_collections.

Train the model using the following command.

export OUTPUT_DIR="toy"
CUDA_VISIBLE_DEVICES=0 accelerate launch --config_file default_config.yaml train_dreambooth_dpg.py \
--pretrained_model_name_or_path path/to/pretrained_stable_diffusion \
--instance_data_dir path/to/personalized_collections \
--instance_prompt "a photo of sks backpack" \
--with_prior_preservation --prior_loss_weight=1.0 \
--class_data_dir="path_class_images_backpack" \
--output_dir=$OUTPUT_DIR \
--class_prompt="a photo of backpack" \
--resolution=512 --train_batch_size=1 --max_train_steps=1000 --learning_rate=1e-6  \
--num_class_images=8 --lr_warmup_steps=0 \
--lr_scheduler="constant" \
--train_text_encoder

Using 'DINO' reward

Download ViT-S/16 ckpt from the official website https://github.com/facebookresearch/dino.

Inference

Use the following command for inference

CUDA_VISIBLE_DEVICES=0 python generate_images.py --ckpt_path /path/to/model --prompt "A sks backpack on the beach"

Visualization Examples

<img width="852" alt="image" src="https://github.com/user-attachments/assets/a1ba1687-8864-4c19-872b-a1fef50c51f6">

Todo

<!-- ## **Citation** -->