tchuanm / iterinv

The official implement of "IterInv: Iterative Inversion for Pixel-Level T2I Models".

IterInv: Iterative Inversion for Pixel-Level T2I Models

(NeurIPS 2023 Workshop on Diffusion Models) & ICME 2024

Environment Setting

0. Our code is based on diffusers-0.19.0
1. Download the dataset from GoogleDrive.
2. create environment.

conda create --name floyd --file environment.yml
conda activate floyd

3. If you want to get a prompt of our own images, you can use BLIP_2.ipynb to get the text prompt.

Reconstruction Image

1. Reconstruct a image based on IterInv.

python end2end_inv.py \
    --input_image 'images/pix2pix-zero/cat/cat_7.png'  \
    --results_folder  'output/all_imgs_inversion_in_prompt_file'  \
    --prompt_str 'a cat' 
    --enable_1 \
    --enable_3for2 \
    --enable_3 \
# or
bash bashes/bash_inv_1img.sh                 

2. Reconstruct multiple images based on IterInv.

python bashes/ours_inv_multi_prompt.py

3. Reconstruct based on DDIM Inversion. Choose stage 1/2/3 to groups what you want.

python bashes/ddim_inv_multi_prompts.py 
# including ddim_stage23_inv.py ddim_failure_stage_3.py  end2end_ddim_inv.py to chooose.

4. Reconstruct based on SDXL.

python SDXL.py
# or in SDXL.ipynb to run it one-by-one step for better development. 

Editing images

1. editing with IterInv + DiffEdit.

python end2end_diffedit.py \ 
    --enable_1  --enable_3for2 --enable_3     \
    --inpaint_strength 0.4 

Quantitative comparison of inversion results.

python evaluation_scores.py  
# or single-step debug in evaluation_scores.ipynb

Change the folder to choose what you want to evaluate.

Acknowledgement

Thanks for the diffusers and DeepFloyd-IF, which helps us to quickly implement our ideas.
Note: this is our draft code release for paper. The cleaned version will be released later.

Citation

If our work is useful for your research, you can consider citing:

@article{tang2023iterinv,
  title={IterInv: Iterative Inversion for Pixel-Level T2I Models},
  author={Tang, Chuanming and Wang, Kai and van de Weijer, Joost},
  journal={arXiv preprint arXiv:2310.19540},
  year={2023}
}