REALEDIT: Reddit Edits As a Large-scale Empirical Dataset for Image Transformations

Project page: https://peter-sushko.github.io/RealEdit/
Data: https://huggingface.co/datasets/peter-sushko/RealEdit

Paper: https://arxiv.org/pdf/2502.03629

There are 2 ways to run inference: either via Diffusers or original InstructPix2Pix pipeline.

Option 1: With 🧨Diffusers:

Install necessary libraries:

pip install torch==2.7.0 diffusers==0.33.1 transformers==4.51.3 accelerate==1.6.0 pillow==11.2.1 requests==2.32.3

Then run:

import torch
import requests
import PIL
from diffusers import StableDiffusionInstructPix2PixPipeline, EulerAncestralDiscreteScheduler

model_id = "peter-sushko/RealEdit"
pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    safety_checker=None
)
pipe.to("cuda")
pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)

url = "https://raw.githubusercontent.com/AyanaBharadwaj/RealEdit/refs/heads/main/example_imgs/simba.jpg"
def download_image(url):
    image = PIL.Image.open(requests.get(url, stream=True).raw)
    image = PIL.ImageOps.exif_transpose(image)
    image = image.convert("RGB")
    return image
image = download_image(url)

prompt = "give him a crown"
result = pipe(prompt, image=image, num_inference_steps=50, image_guidance_scale=2).images[0]
result.save("output.png")

Option 2: via InstructPix2Pix pipeline:

Clone the repository and set up the directory structure:

git clone https://github.com/timothybrooks/instruct-pix2pix.git
cd instruct-pix2pix
mkdir checkpoints

Download the fine-tuned checkpoint into the checkpoints directory:

cd checkpoints
# wget https://huggingface.co/peter-sushko/RealEdit/resolve/main/realedit_model.ckpt

Return to the repo root and follow the InstructPix2Pix installation guide to set up the environment.

Edit a single image

python edit_cli.py \
  --input [YOUR_IMG_PATH] \
  --output imgs/output.jpg \
  --edit "YOUR EDIT INSTRUCTION" \
  --ckpt checkpoints/realedit_model.ckpt

Citation

If you find this checkpoint helpful, please cite:

@misc{sushko2025realeditredditeditslargescale,
      title={REALEDIT: Reddit Edits As a Large-scale Empirical Dataset for Image Transformations}, 
      author={Peter Sushko and Ayana Bharadwaj and Zhi Yang Lim and Vasily Ilin and Ben Caffee and Dongping Chen and Mohammadreza Salehi and Cheng-Yu Hsieh and Ranjay Krishna},
      year={2025},
      eprint={2502.03629},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2502.03629}, 
}