At the core of InfiniteYou is the InfuseNet module, a specialized neural component that injects identity features into the diffusion process through residual connections. This approach ensures that generated images retain a strong resemblance to the original subject, even as backgrounds, outfits, and poses are dramatically altered according to user prompts. The model employs a multi-stage training strategy, including pretraining and supervised fine-tuning with synthetic single-person multi-sample data, which enhances text-image alignment and overall image aesthetics. InfiniteYou offers two model variants: 'aes_stage2' for improved text-image alignment and visual appeal, and 'sim_stage1' for maximum facial similarity, giving users flexibility based on their priorities.
InfiniteYou is engineered with a plug-and-play design, making it compatible with a wide range of existing image generation methods, tools, and extensions such as ControlNets and LoRAs. The framework supports optional control images for pose or structure guidance, and provides adjustable parameters for conditioning scale and guidance, allowing for fine-tuned customization. Open-sourced for research and creative use, InfiniteYou stands out for its ability to generate high-fidelity, identity-consistent images with superior realism and prompt alignment, outperforming previous solutions in both quality and flexibility.
Key features include:
- Preserves personal identity features in all generated images
- Text-prompt-based customization for flexible scene and style changes
- InfuseNet module for advanced identity feature injection
- Multi-stage training for improved text-image alignment and aesthetics
- Two model variants: aes_stage2 (aesthetics) and sim_stage1 (similarity)
- Plug-and-play compatibility with ControlNets, LoRAs, and other tools
- Supports optional control images for pose or structure
- Open-source code and models for research and creative applications