Abstract: Achieving flexible and high-fidelity identity-preserved image generation remains formidable, particularly with advanced Diffusion Transformers (DiTs) like FLUX. We introduce InfiniteYou (InfU), one of the earliest robust frameworks leveraging DiTs for this task. InfU addresses significant issues of existing methods, such as insufficient identity similarity, poor text-image alignment, and low generation quality and aesthetics. Central to InfU is InfuseNet, a component that injects identity features into the DiT base model via residual connections, enhancing identity similarity while maintaining generation capabilities. A multi-stage training strategy, including pretraining and supervised fine-tuning (SFT) with synthetic single-person-multiple-sample (SPMS) data, further improves text-image alignment, ameliorates image quality, and alleviates face copy-pasting. Extensive experiments demonstrate that InfU achieves state-of-the-art performance, surpassing existing baselines. In addition, the plug-and-play design of InfU ensures compatibility with various existing methods, offering a valuable contribution to the broader community.
We released two model variants of InfiniteYou-FLUX v1.0: aes_stage2 and sim_stage1. The aes_stage2 is our model after stage-2 SFT, which is used by default for better text-image alignment and aesthetics. If you wish to achieve higher ID similarity, please try sim_stage1.
To better fit specific personal needs, we find that two arguments are highly useful to adjust in our code: --infusenet_conditioning_scale (default: 1.0) and --infusenet_guidance_start (default: 0.0). Usually, you may NOT need to adjust them. If necessary, start by trying a slightly larger --infusenet_guidance_start (e.g., 0.1) only (especially helpful for sim_stage1). If still not satisfactory, then try a slightly smaller --infusenet_conditioning_scale (e.g., 0.9).
We also provided two LoRAs (Realism and Anti-blur) to enable additional usage flexibility. They are entirely optional, which are examples to facilitate users to try but are NOT used in our paper.
If the generated gender is not preferred, try adding specific words in the text prompt, such as 'a man', 'a woman', etc. We encourage using inclusive and respectful language.
Stage-1 model before SFT. Higher identity similarity.
🆚 Comparison with State-of-the-Art Relevant Methods
comparative_results
Qualitative comparison results of InfU with the state-of-the-art baselines, FLUX.1-dev IP-Adapter and PuLID-FLUX. The identity similarity and text-image alignment of the results generated by FLUX.1-dev IP-Adapter (IPA) are inadequate. PuLID-FLUX generates images with decent identity similarity. However, it suffers from poor text-image alignment (Columns 1, 2, 4), and the image quality (e.g., bad hands in Column 5) and aesthetic appeal are degraded. In addition, the face copy-paste issue of PuLID-FLUX is evident (Column 5). In comparison, the proposed InfU outperforms the baselines across all dimensions.
⚙️ Plug-and-Play Property with Off-the-Shelf Popular Approaches
plug_and_play
InfU features a desirable plug-and-play design, compatible with many existing methods. It naturally supports base model replacement with any variants of FLUX.1-dev, such as FLUX.1-schnell for more efficient generation (e.g., in 4 steps). The compatibility with ControlNets and LoRAs provides more controllability and flexibility for customized tasks. Notably, the compatibility with OminiControl extends our potential for multi-concept personalization, such as interacted identity (ID) and object personalized generation. InfU is also compatible with IP-Adapter (IPA) for stylization of personalized images, producing decent results when injecting style references via IPA. Our plug-and-play feature may extend to even more approaches, providing valuable contributions to the broader community.
📜 Disclaimer and Licenses
Most images used in this repository and related demos are sourced from consented subjects, with a few taken from public domains or generated by the models. These pictures are intended solely to showcase the capabilities of our research. If you have any concerns, please feel free to contact us, and we will promptly remove any inappropriate content.
This research aims to positively impact the field of Generative AI. Users are granted the freedom to create images using this tool, but they must comply with local laws and use it responsibly. The developers do not assume any responsibility for potential misuse by users.
📖 Citation
If you find InfiniteYou useful for your research or applications, please cite our paper:
bibtex
1@article{jiang2025infiniteyou,
2 title={{InfiniteYou}: Flexible Photo Recrafting While Preserving Your Identity},
3 author={Jiang, Liming and Yan, Qing and Jia, Yumin and Liu, Zichuan and Kang, Hao and Lu, Xin},
4 journal={arXiv preprint},
5 volume={arXiv:2503.16418},
6 year={2025}
7}
We also appreciate it if you could give a star ⭐ to our Github repository. Thanks a lot!
