Perfusion
_{Key-Locked Rank One Editing for Text-to-Image Personalization}

A Pytorch implementation of the paper Key-Locked Rank One Editing for Text-to-Image Personalization (project page).

⏳ To do

• Multiple concepts inference
• Support SDXL-1.0
• CLIP metrics
• Evaluation

News

• Support advanced sampler from Stable Diffusion, like EulerEDMSampler.
• Now we support training and generating using SD V2.1 as the basement!

Samples

Paper

Our

Using CLIP similarity to automatically select a balanced weight is necessary. We'll implement it in the near future.

Environment

Create and activate the conda environment:

conda env create -f environment.yaml
conda activate perfusion

Training

Download the SD V1.5 or SD V2.1 to ./ckpt/.

Then run the commands:

python main.py \
    --name teddy \
    --base ./configs/perfusion_teddy.yaml \
    --basedir ./ckpt \
    -t True \
    --gpus 0,

or:

python main.py \
    --name teddy \
    --base ./configs/perfusion_teddy_sd_v2.yaml \
    --basedir ./ckpt \
    -t True \
    --gpus 0,

To prepare your own training data, please ensure that they are placed in a folder /path/to/your/images/. You need to download pretrained weight of clipseg:

wget https://owncloud.gwdg.de/index.php/s/ioHbRzFx6th32hn/download -O ./clipseg/weights.zip
unzip -d ./clipseg/weights -j ./clipseg/weights.zip

Then run:

python ./data/soft_segment.py --image_dir /path/to/your/images/ --super_class your_own_super_class

Modify the initializer_words, data_root, flip_p in ./configs/perfusion_custom.yaml or ./configs/perfusion_custom_sd_v2.yaml.

Finally, run:

python main.py \
    --name experiment_name \
    --base ./configs/perfusion_custom.yaml \
    --basedir ./ckpt \
    -t True \
    --gpus 0,

You can find weights along with tensorboard in ./ckpt.

Pretrained Weight

Pretrained concepts' weights can be found in ./ckpt.

Generating

Personalized samples can be obtained by running the command

python scripts/perfusion_txt2img.py --ddim_eta 0.0 \
                                    --steps 50  \
                                    --scale 6.0 \
                                    --beta 0.7 \
                                    --tau 0.15 \
                                    --n_samples 4 \
                                    --n_iter 1 \
                                    --personalized_ckpt ./ckpt/teddy.ckpt \
                                    --prompt "photo of a {}"

Global Locking

Global locking will be applied with the label --global_locking.

Stable Diffusion V2

Set --config configs/perfusion_inference_sd_v2.yaml and --ckpt ./ckpt/v2-1_512-ema-pruned.ckpt when using SD v2.1.

Multiple Concepts Generating

When generating with multiple concepts, use commas to separate checkpoints like --personalized_ckpt /path/to/personalized/ckpt1,/path/to/personalized/ckpt2, and use {1},{2}, ..., {n} to distinguish different concepts in the prompt as --prompt "photo of a {1} and {2}".

If you want to apply different biases and temperatures to each concept, set --beta b1,b2 and --tau t1,t2.

Advanced Sampler

If you want to use advanced samplers other than DDIM, use the label --advanced_sampler. The default advanced sampler is the EulerEDMSampler. You can modify ./configs/sampler/sampler.yaml and ./configs/denoiser/denoiser.yaml depend on you preference.

Following are results by utilizing SD v2.1 and EulerEDMSampler.

BibTeX

If you find this repository useful, please cite origin papers using the following.

@article{Tewel2023KeyLockedRO,
    title   = {Key-Locked Rank One Editing for Text-to-Image Personalization},
    author  = {Yoad Tewel and Rinon Gal and Gal Chechik and Yuval Atzmon},
    journal = {ACM SIGGRAPH 2023 Conference Proceedings},
    year    = {2023},
    url     = {https://api.semanticscholar.org/CorpusID:258436985}
}

@inproceedings{Meng2022LocatingAE,
    title   = {Locating and Editing Factual Associations in GPT},
    author  = {Kevin Meng and David Bau and Alex Andonian and Yonatan Belinkov},
    booktitle = {Neural Information Processing Systems},
    year    = {2022},
    url     = {https://api.semanticscholar.org/CorpusID:255825985}
}