Dataset, code, and model. For more information, read our paper which was accepted to CVPR 2025 or check out our project page!
- Instructions for finetuning InstructPix2Pix with our config file
- Instructions for evaluating on LLM-based metrics and Elo scores
You can access our dataset on Huggingface. We provide a training set of 48,002 editing requests with 1-5 ground truth outputs and a test set of 9,337 editing requests with 1 manually verified ground truth output.
Our model checkpoint is available on Huggingface.
You can load the model using the diffusers library.
We provide code to make a single edit in edit_single_image.ipynb. We provide code to create generations from a csv (containing columns input_image_name and instruction) and image directory in inference.py.
To run inference.py, use the following command:
python inference.py --csv_path path/to/your/csv --image_dir path/to/your/image/folder --output_dir path/to/your/output/folder [--num_inference_steps NUM_STEPS] [--image_guadance_scale IMG_SCALE] [--text_guidance_scale TEXT_SCALE]
We have provided the code for some automatic evaluations in metrics_calculation.py.
If you found our code helpful, please cite our paper!
@article{sushko2024realedit,
title = {REALEDIT: Reddit Edits As a Large-scale Empirical Dataset for Image Transformations},
author = {Sushko, Peter and Bharadwaj, Ayana and Lim, Zhi Yang and Ilin, Vasily and Caffee, Ben and Chen, Dongping and Salehi, Mohammadreza and Hsieh, Cheng-Yu and Krishna, Ranjay},
journal = {arXiv preprint},
year = {2024},
}