ByteDance's InfiniteYou lets users generate unlimited variations of portrait photos

ByteDance has developed a new approach to AI portrait generation that tackles common problems like inconsistent facial features and poor prompt following.

Unlike previous solutions such as PuLID-FLUX that directly modify AI model attention, InfuseNet processes facial features as a parallel information layer. This keeps the core AI model intact while improving portrait generation quality.

InfuseNet processes identity features and control inputs separately from the main AI model, allowing for better preservation of facial characteristics while following text prompts. | Image: ByteDance

The system uses a two-stage training process. It starts with real portrait photos before moving to synthetic images, which are generated by the system using specialized optimization modules.

The approach significantly improves output quality, according to ByteDance. Generated images maintain better similarity to the original person and follow text prompts more accurately. It also avoids common problems like direct face copying and quality degradation that plague other approaches.

Side-by-side comparison shows InfiniteYou (right) maintaining better facial consistency and following text prompts more accurately than previous approaches. | Image: ByteDance

In user testing with 16 participants, InfiniteYou showed clear advantages over existing solutions. When evaluating facial similarity, text prompt accuracy, image quality, and aesthetics, 72.8 percent preferred InfiniteYou's results compared to 27.2 percent for PuLID-FLUX.

InfiniteYou works with popular AI tools including ControlNet and LoRA. The system needs just four processing steps to generate images and lets users modify both people and objects within them. While acknowledging these improvements, ByteDance notes there's still room to enhance facial similarity and image quality.

The system integrates with existing AI tools like ControlNet and LoRA, enabling a wide range of portrait modifications and style adjustments. | Bild: ByteDance

Where to try InfiniteYou

ByteDance has released InfiniteYou's code on GitHub and made the model weights available through Hugging Face. An experimental demo is planned but hasn't launched yet.

As a member of the Content Authenticity Initiative, ByteDance works with camera manufacturers and media providers to implement C2PA metadata for identifying AI-generated content. However, the company hasn't specified whether InfiniteYou will include these authentication features.

The release follows ByteDance's recent AI developments, including OmniHuman-1 for photorealistic animations and the Goku series for AI avatar videos.