PixelHacker: Image Inpainting with Structural and Semantic Consistency

SOTA performance on Places2, CelebA-HQ, and FFHQ & Superior structural and semantic consistency

Ziyang Xu¹, Kangsheng Duan¹, Xiaolei Shen², Zhifeng Ding², Wenyu Liu¹, Xiaohu Ruan²,
Xiaoxin Chen², Xinggang Wang^{1 📧}

(^📧) Corresponding Author.

¹ Huazhong University of Science and Technology. ² VIVO AI Lab.

🌟 Highlights

• Latent Categories Guidance (LCG): Simple yet effective inpainting paradigm with superior structural and semantic consistency. Let's advance inpainting research to challenge more complex scenarios!
• PixelHacker: Diffusion-based inpainting model trained with LCG, outperforming SOTA performance across multiple natural-scene (Places2) and portrait (CelebA-HQ, and FFHQ) benchmarks!
• Comprehensive SOTA Performance：
  • Places2 (Natural Scene)
    • Evaluated at 512 resolution using 10k test set images with 40-50% masked regions, PixelHacker achieved the best performance with FID 8.59 and LPIPS 0.2026.
    • Evaluated at 512 resolution using 36.5k validation set images with large and small mask settings, PixelHacker achieved the best performance on FID (large: 2.05, small: 0.82) and U-IDS (large:36.07, small:42.21), and the second best performance on LPIPS (large:0.169, small:0.088).
    • Evaluated at 256 and 512 resolutions using validation set images with a highly randomised masking strategy, PixelHacker achieved the best performance at 512 resolution with FID 5.75 and LPIPS 0.305, and the second best performance at 256 resolution with FID 9.25 and LPIPS 0.367.
  • CelebA-HQ (Portrait Scene)
    • Evaluated at 512 resolution, PixelHacker achieved the best performance with FID 4.75 and LPIPS 0.115.
  • FFHQ (Portrait Scene)
    • Evaluated at 256 resolution, PixelHacker achieved the best performance with FID 6.35 and LPIPS 0.229.

🔥 News

• June 18, 2026: 🎉 Following up on PixelHacker, our latest AI Image Inpainting work, Moebius, has been accepted by ECCV'26. We challenge the "scale-at-all-costs" path dependency dictated by the Scaling Law narrative and provide a core insight into breaking the impossible triangle of "low parameters, fast inference, and high quality".

• June 01, 2026: 📰✨ The journal version of PixelHacker is coming soon. Stay tuned! ☕ Next, we will gradually update the README and project page.

• May 20, 2025: 🔥 We have released the code and weights. The weights include the pretrained and all fine-tuned versions, each only 0.8B params. Feel free to play!

• May 5, 2025: 🎉 PixelHacker has achieved the No. 1 daily ranking on Hugging Face!

• May 1, 2025: 🔥 We have released the project page with 63+ demos on natural and portrait scenes. Have fun! 🤗

• April 30, 2025: 🔥 We have released the arXiv paper for PixelHacker. The code and project page will be released soon.

🏕️ Performance on Natural Scene

🤗 Performance on Portrait Scene

📦 Environment Setups

• torch 2.3.0
• transformers 4.40.0
• diffusers 0.30.2
• See 'requirements.txt' for detailed Python libraries required

conda create -n pixelhacker python=3.10
conda activate pixelhacker
# cd /xx/xx/PixelHacker
pip install -r requirements.txt

🗃️ Model Checkpoints

• Download the checkpoint of VAE and put it into ../PixelHacker/vae.

• Download the checkpoints of pretrained version, fine-tuned version (places2), fine-tuned version (celeba-hq), fine-tuned version (ffhq), and put them into ../PixelHacker/weight.

• Finally, the detailed organizational form is as follows:

├── PixelHacker
│    ├── weight
│        ├── pretrained
│            ├── diffusion_pytorch_model.bin
│        ├── ft_places2
│            ├── diffusion_pytorch_model.bin
│        ├── ft_celebahq
│            ├── diffusion_pytorch_model.bin
│        ├── ft_ffhq
│            ├── diffusion_pytorch_model.bin
│    ├── vae
│        ├── config.json
│        ├── diffusion_pytorch_model.bin
│    ├── ...

🔮 Inference

You can run the following code directly to get the inpainting result of the example image-mask pair, and the result will be generated in ../PixelHacker/outputs. If you want to infer on custom data, just place the image and mask with the same name in ../PixelHacker/imgs and ../PixelHacker/masks, respectively, then run the following code as well.

python infer_pixelhacker.py \
--config config/PixelHacker_sdvae_f8d4.yaml \
--weight weight/ft_places/diffusion_pytorch_model.bin

🎓 Citation

@misc{xu2025pixelhacker,
      title={PixelHacker: Image Inpainting with Structural and Semantic Consistency}, 
      author={Ziyang Xu and Kangsheng Duan and Xiaolei Shen and Zhifeng Ding and Wenyu Liu and Xiaohu Ruan and Xiaoxin Chen and Xinggang Wang},
      year={2025},
      eprint={2504.20438},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.20438}, 
}