Skip to content

Latest commit

 

History

History
193 lines (158 loc) · 7.6 KB

README.md

File metadata and controls

193 lines (158 loc) · 7.6 KB
Raw

Cones 2

Official repo for Cones 2: Customizable Image Synthesis with Multiple Subjects | Project Page.

Cones 2 allows you to represent a specific subject as a residual embedding by fine-tuning text encoder in a pre-trained text-to-image diffusion model, such as Stable Diffusion. After tuning, we only need to save the residual between tuned text-encoder and frozen one. Thus, the storage space required for each additional subject is only 5 KB. This step only takes about 20~30 minutes on a single 80G A100 GPU for each subject.

When sampling, our layout guidance sampling method further allows you to employ a easy-to-obtain layout as guidance for multiple subjects arrangement as shown in the following figure.

Results

All results are synthesized by pre-trained Stable Diffusion v2.1 models and our customized residual embeddings. We show diverse results on various categories of images, including scene, pet, personal toy, human etc.. For more results, please refer to our paper or website.

Two-Subject Results

two subject cases

Three-Subject Results

three subject cases

Four-Subject Results

four subject cases

More Results

more challenging cases

Method

method

(a) Given few-shot images of the customized subject, we fine-tune the text encoder to learn a residual embedding on top of the base embedding of raw subject. (b) Based on the residual embeddings, we then propose to employ layout as the spatial guidance for subject arrangement into the attention maps. After that, we could strengthen the signal of target subjects and weaken the signal of irrelevant subjects. For more details, please refer to our paper.

Getting Started

Installing the dependencies

The implementation of Cones 2 is based entirely on the diffuser. Before running out code, make sure to install the library's training dependencies. To do this, execute the following steps in a new virtual environment:

git clone https://github.com/damo-vilab/Cones-V2.git
cd Cones-V2
pip install -r requirements.txt

Then initialize an 🤗—Accelerate environment with:

accelerate config

Or for a default accelerate configuration without answering questions about your environment

accelerate config default

Training (Flower example)

Firstly, let's download dataset from here and unzip it to ./data. Now a few images of Flower (download to ./data/flower) are used to learn its customized residual embedding.

export MODEL_NAME='path-to-stable-diffusion-v2-1'
export INSTANCE_DIR="./data/flower"
export OUTPUT_DIR="path-to-save-model"
accelerate launch train_cones2.py \
  --pretrained_model_name_or_path=$MODEL_NAME  \
  --instance_data_dir=$INSTANCE_DIR \
  --instance_prompt="flower" \
  --token_num=1 \
  --output_dir=$OUTPUT_DIR \
  --resolution=768 \
  --train_batch_size=1 \
  --gradient_accumulation_steps=1 \
  --learning_rate=5e-6 \
  --lr_scheduler="constant" \
  --lr_warmup_steps=0 \
  --max_train_steps=4000 \
  --loss_rate_first=1e-2 \
  --loss_rate_second=1e-3

Inference

Once you have trained several residual embeddings of different subjects using the above command, you can run our layout guidance sampling method simply using inference.py. We provide several pre-trained models for quick validation.

residual embedding raw token download
barn barn barn.pt
white dog dog dog.pt
flower flower flower.pt
lake lake lake.pt
mug mug mug.pt
sunglasses sunglasses sunglasses.pt
bag bag bag.pt
robot robot toy robot_toy.pt

Remember to provide a pre-defined layout like layout_example.png and a json file with the info about the details of the inference settings. The json file should include the following information:

  • "prompt": the text prompt you want to generate.
  • "residual_dict": the paths to all the required residual embeddings.
  • "color_context": the color information of different regions in the layout and their corresponding subjects, along with the weight for strengthening the signal of target subject. (default: 2.5).
  • "guidance_steps": the number of steps of the layout guidance.
  • "guidance_weight": the strength of the layout guidance (default: 0.08, we recommond 0.05 ~ 0.10).
  • "weight_negative": the weight for weakening the signal of irrelevant subject.
  • "layout": the path to user-defined layout image.
  • "subject_list": the list containing all the subjects to be customized and their corresponding positions in the prompt.

Then you can simply run inference script with:

python inference.py --pretrained_model_name_or_path /path/to/stable-diffusion-2-1 --inference_config guidance_config_example.json

References

@article{liu2023cones,
  title={Cones 2: Customizable Image Synthesis with Multiple Subjects},
  author={Liu, Zhiheng and Zhang, Yifei and Shen, Yujun and Zheng, Kecheng and Zhu, Kai and Feng, Ruili and Liu, Yu and Zhao, Deli and Zhou, Jingren and Cao, Yang},
  journal={arXiv preprint arXiv:2305.19327},
  year={2023}
}

Acknowledgements

We thank Stable Diffusion v2.1 and diffuser for providing pre-trained model and an open-source codebase.