UltraEdit

This repository contains code, models, and datasets for UltraEdit.

Introduction

UltraEdit, a large-scale (~4M editing samples), automatically generated dataset for instruction-based image editing. Our key idea is to address the drawbacks in existing image editing datasets like InstructPix2Pix and MagicBrush, and provide a systematic approach to producing massive and high-quality image editing samples.

UltraEdit offers several distinct advantages:

1. It features a broader range of editing instructions by leveraging the creativity of large language models (LLMs) alongside in-context editing examples from human raters.
2. Its data sources are based on real images, including photographs and artworks, which provide greater diversity and reduced bias compared to datasets solely generated by text-to-image models.
3. It also supports region-based editing, enhanced by high-quality, automatically produced region annotations.

Our experiments show that canonical diffusion-based editing baselines trained on UltraEdit set new records on various benchmarks. Our analysis further confirms the crucial role of real image anchors and region-based editing data.

Training

Setup:

pip install -r requirements

cd diffusers && pip install -e .

Training with stable-diffusion3

Stage 1: Free-form image editing

bash scripts/run_sft_512_sd3_stage1.sh

Stage 2: Mix training

bash scripts/run_sft_512_with_mask_sd3_stage2.sh

Training with stable-diffusion-xl

Stage 1: Free-form image editing

bash scripts/run_sft_512_sdxl_stage1.sh

//: # ()

Training with stable-diffusion1.5

Stage 1: Free-form image editing

bash scripts/run_sft_512_sd15_stage1.sh

Stage 2: Mix training

bash scripts/run_sft_512_with_mask_sd15_stage2.sh

Example

Below is an example of how to use our pipeline for image editing. Given an input image and a mask image, the model can generate the edited result according to the provided prompt.

# For Editing with SD3
import torch
from diffusers import StableDiffusion3InstructPix2PixPipeline
from diffusers.utils import load_image
import requests
import PIL.Image
import PIL.ImageOps
pipe = StableDiffusion3InstructPix2PixPipeline.from_pretrained("BleachNick/SD3_UltraEdit_w_mask", torch_dtype=torch.float16)
pipe = pipe.to("cuda")
prompt="What if the horse wears a hat?"
img = load_image("input.png").resize((512, 512))
mask_img = load_image("mask_img.png").resize(img.size)
# For free form Editing, seed a blank mask
# mask_img = PIL.Image.new("RGB", img.size, (255, 255, 255))
image = pipe(
    prompt,
    image=img,
    mask_img=mask_img,
    negative_prompt="",
    num_inference_steps=50,
    image_guidance_scale=1.5,
    guidance_scale=7.5,
).images[0]
image.save("edited_image.png")
# display image

Citation

If you find our work useful, please kindly cite

@misc{zhao2024ultraeditinstructionbasedfinegrainedimage,
      title={UltraEdit: Instruction-based Fine-Grained Image Editing at Scale}, 
      author={Haozhe Zhao and Xiaojian Ma and Liang Chen and Shuzheng Si and Rujie Wu and Kaikai An and Peiyu Yu and Minjia Zhang and Qing Li and Baobao Chang},
      year={2024},
      eprint={2407.05282},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2407.05282}, 
}

//: # ()

//: # ()

//: # ()

//: # ()