[NeurIPS 25] Official implementation of AutoEdit: Automatic Hyperparameter Tuning for Image Editing

This is the official implementation of AutoEdit: Automatic Hyperparameter Tuning for Image Editing (NeurIPS 2025). The paper can be found here:

To-do list

• Release inference code.
• Release training code.

Installation

For editing methods running with SD 1.4, please refer to the environment installation in python_env/environment.yml:

conda env export -n p2p > python_env/environment.yml

or you can install the environment by pip on your own. Note the version of some important libraries:

accelerate==1.4.0
diffusers==0.12.0
torch==2.1.2
transformers==4.37.2

Note that the version of diffusers is low with the SD 1.4 model, I will try to update the code to make it compatible with recent diffusers version.

Stay tuned for SDXL and DiT model.

Checkpoint

Checkpoint path can be downloaded here. Each folder will be the checkpoint of AutoEdit with the corresponding method (see the Running AutoEdit section).

Running AutoEdit

You can run AutoEdit by following these scripts:

DDPM-Inversion: Search for inversion timestep:

python run_training_wo_attention.py \
    --checkpoint "exp/ddpm_ppo/checkpoint.ckpt" \
    --input_image_path "assets/cake.jpg" \
    --src_prompt "a round cake with orange frosting on a wooden plate" \
    --tgt_prompt "a square cake with orange frosting on a wooden plate" \
    --save_edit_path "output/edit_cake.png"

P2P: Conduct the DDPM Inversion and cross-attention ratio search:

python run_training.py \
  --checkpoint "exp/ddpm_ppo/checkpoint.ckpt" \
  --input_image_path "assets/cake.jpg" \
  --src_prompt "a round cake with orange frosting on a wooden plate" \
  --tgt_prompt "a square cake with orange frosting on a wooden plate" \
  --save_edit_path "output/edit_cake.png"

Null-text: Null-text inversion and searching for inversion timestep

python null_text_inversion.py \
  --checkpoint_path "exp/null_text_ppo/checkpoint.ckpt" \
  --input_image_path "assets/cake.jpg" \
  --src_prompt "a round cake with orange frosting on a wooden plate" \
  --tgt_prompt "a square cake with orange frosting on a wooden plate" \
  --save_edit_path "output/edit_cake.png"

Adaptive CFG:

python run_training_cfg.py \
    --checkpoint "exp/cfg_ppo/checkpoint.ckpt" \
    --input_image_path "assets/cake.jpg" \
    --src_prompt "a round cake with orange frosting on a wooden plate" \
    --tgt_prompt "a square cake with orange frosting on a wooden plate" \
    --save_edit_path "output/edit_cake.png"

All scripts are available at scripts folder.

Training AutoEdit

Prepare the dataset: You can find the dataset here. Put it into the data folder.

Training AutoEdit for DDPM inversion:

Stage 1:SFT training

python run_training_wo_attention.py \
    --exp_name "exp/ddpm_ppo_sft" \
    --num_epochs 5 \
    --low 0.05 \
    --range 0.6 \
    --annotation_folder "data/EditBench/EditData" \
    --train_sft

Stage 2: PPO training

python run_training_wo_attention.py \
    --exp_name "exp/ddpm_ppo_sft" \
    --num_epochs 15 \
    --annotation_folder "data/EditBench/EditData" \
    --train_ppo \
    --checkpoint "path/to/checkpoint_sft"

Citation

If you find our work interesting, please considering cite:

@article{pham2025autoedit,
  title={AutoEdit: Automatic Hyperparameter Tuning for Image Editing},
  author={Pham, Chau and Dao, Quan and Bhosale, Mahesh and Tian, Yunjie and Metaxas, Dimitris and Doermann, David},
  journal={arXiv preprint arXiv:2509.15031},
  year={2025}
}