---
license: apache-2.0
extra_gated_fields:
  Name: text
  Institute: text
  Institutional Email: text
  I agree to use this model for non-commercial use ONLY: checkbox
---
<p align="center">
  <h1 align="center">Towards Enhanced Image Inpainting:<br>
Mitigating Unwanted Object Insertion and Preserving Color Consistency</h1>
<center>Yikai Wang*, Chenjie Cao*, Junqiu Yu*, Ke Fan, Xiangyang Xue, Yanwei Fu†.<br>
Fudan University<br>
<b>CVPR 2025 <font color="#ed7748">(Highlight)</font></b>
</center>
  <p align="center">
    <a href="https://arxiv.org/abs/2312.04831"><img alt='arXiv' src="https://img.shields.io/badge/arXiv-2312.04831-b31b1b.svg"></a>
    <a href="https://yikai-wang.github.io/asuka/"><img alt='page' src="https://img.shields.io/badge/Project-Website-orange"></a>

  </p>
  <br>
</p>

## Overview
This repo contains the proposed ASUKA model in our paper "[Towards Enhanced Image Inpainting: Mitigating Unwanted Object Insertion and Preserving Color Consistency](https://arxiv.org/abs/2312.04831)".

> ASUKA solves two issues existed in current diffusion and rectified flow inpainting models:
<b>Unwanted object insertion</b>, where randomly elements that are not aligned with the unmasked region are generated;
<b>Color-inconsistenc</b>y, the color shift of the generated masked region, causing smear-like traces.
ASUKA proposes a post-training procedure for these models, significantly mitigates object hallucination and improves color consistency of inpainted results.

We released ASUKA for [FLUX.1-Fill-dev](https://huggingface.co/black-forest-labs/FLUX.1-Fill-dev), denoted as ASUKA(FLUX.1-Fill).
The code and dataset can be found at [here](https://github.com/Yikai-Wang/asuka-misato).
We are actively working to improve both our model and evaluation dataset.
If you encounter failure cases with ASUKA (FLUX.1-Fill) or have challenging examples in image inpainting, we would love to hear from you. Please email them to yi-kai.wang@outlook.com.
We truly appreciate your contributions!

## Modifications to FLUX
- The text conditional input of CLIP and T5 is replaced by the MAE condition to mitigate object hallucination.
- The decoder is replaced by our conditional decoder to enhance color consistency.

## BibTeX
If you find our repo helpful, please consider cite our paper :)
```bibtex
@inproceedings{wang2025towards,
  title={Towards Enhanced Image Inpainting: Mitigating Unwanted Object Insertion and Preserving Color Consistency.},
  author={Wang, Yikai and Cao, Chenjie and Yu, Junqiu and Fan, Ke and Xue, Xiangyang and Fu, Yanwei},
  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
  year={2025}
}
```