WithAnyone - Fudan joint step leap star open source AI photo generation model

What's WithAnyone?

WithAnyone is an AI photo generation model jointly developed by Fudan University and StepStar, which solves the common "copy and paste" problem in traditional AI image generation and realizes more natural and controllable multi-person image generation. WithAnyone is based on the diffusion model architecture, which effectively reduces artifacts in the generated images and improves identity similarity and diversity. The model has been open-sourced, and the model, dataset, and demo are available on Hugging Face. Users can easily upload their personal photos to generate group photos with any person in multiple scenarios, and experience natural and realistic AI image generation effects.

WithAnyone's main features

• Multi-Photo Generation: The ability to naturally blend multiple people into a single group photo, generating images that are unobtrusive.
• Identity Consistency Maintenance: The identity of the character can be maintained while generating the image, ensuring that the generated image is highly similar to the input character.
• Flexibility to adjust details: Users can flexibly adjust the details of the character's expression, posture, and hairstyle to achieve personalized image generation.
• Reducing "copy and paste" artifactsThe AI image generation process is based on advanced technology, which effectively reduces the common "copy and paste" phenomenon in traditional AI image generation and improves image quality.
• Open Source and Ease of Use: The model is open source and users can find resources on Hugging Face to easily use and experience the technology.

WithAnyone's technical principles

• Large-scale dataset support: The MultiID-2M dataset is used, which contains 500,000 multi-person group photos and a large number of reference images covering a wide range of expressions, hairstyles and angles, providing a rich data base for model training.
• Contrasting identity loss training: Balancing identity fidelity and generative diversity through contrastive identity loss (CIL) and pairwise data training to ensure that the generated images maintain the identity characteristics of the characters while having a natural visual effect.
• Diffusion Model ArchitectureThe diffusion model-based architecture effectively reduces "copy and paste" artifacts while maintaining high identity similarity and improving the overall quality of the generated images.
• identity code: Encoding the character's identity characteristics ensures that the unique attributes of the character can be accurately restored during the generation process to avoid identity confusion.
• Multitasking optimization: Simultaneously optimize multiple tasks, such as identity retention, posture adjustment and background fusion, during training to improve the comprehensive performance and adaptability of the model.

WithAnyone's program address

• Project website:: https://doby-xu.github.io/WithAnyone/
• Github repository:: https://github.com/Doby-Xu/WithAnyone
• HuggingFace Model Library:: https://huggingface.co/WithAnyone/WithAnyone
• arXiv Technical Paper:: https://arxiv.org/pdf/2510.14975
• Online Experience Demo:: https://huggingface.co/spaces/WithAnyone/WithAnyone_demo

Who WithAnyone is for

• social media user: It is hoped that AI technology will be used to generate creative group photos with idols, friends or family members to share on social platforms.
• content creator: Need to quickly generate high-quality group photos of multiple people for creation of videos, articles or advertisements.
• Designer & Photographer: for creative design or virtual shoots to explore new forms of visual expression.
• regular user: Experience the fun of AI technology and create personalized photos for entertainment and remembrance needs.
• Technology Enthusiasts & Developers: Research on AI image generation techniques, learning and secondary development using open source models.