3 weeks ago
12
Researchers at ByteDance Intelligent Creation have developed a new AI model that generates multiple versions of an identity along with its paper, demo, and code. Liming Jiang, a senior research scientist at ByteDance, made the announcement on X on Sunday.
The new AI model called InfiniteYou (InfU) aims to address the challenges of identity-preserved image generation. One can create multiple versions of their identity in different settings by using prompts as required, ensuring good accuracy. The model leverages Diffusion Transformers (DiTs) to generate images that not only maintain the identity of a person from a source photograph but also allow for flexible text-based editing.
InfU aims to overcome the limitations found in existing methods, such as insufficient identity similarity, poor text-image alignment, and low generation quality. The core of InfU is InfuseNet, a component designed to inject identity features into the DiT base model through residual connections. This process enhances identity similarity while preserving the model’s generative capabilities.
To further refine the model’s performance, a multi-stage training strategy was employed, incorporating pretraining and supervised fine-tuning (SFT) with synthetic single-person-multiple-sample (SPMS) data. The training approach was designed to improve text-image alignment, enhance image quality, and mitigate face copy-pasting issues.
The official website mentioned, “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.”
“The compatibility with ControlNets and LoRAs provides more controllability and flexibility for customised tasks. Notably, the compatibility with OminiControl extends our potential for multi-concept personalisation, such as interacted identity (ID) and object personalised generation,” the paper added.
The code is available on the GitHub page, and one can access the demo and the model on Hugging Face to try it out.
ByteDance has been making several developments in 2025, including Goku as an alternative to Google’s Luma and a React Native killer. The AI model adds to its list of exciting developments so far.
