InfiniteYou

InfiniteYou (InfU), from ByteDance Intelligent Creation (ICCV 2025 Highlight), tackles identity-preserved generation with diffusion transformers like FLUX. The core is InfuseNet, a component that injects identity features from a reference face into the DiT base model through residual connections, so the output keeps the person's likeness while following the text prompt for scene, style, and pose. A multi-stage training strategy with synthetic single-person-multiple-sample data improves text-image alignment and, the authors claim, reduces the 'face pasting' artifacts that plague identity methods. Two checkpoints ship: aes_stage2 for better aesthetics and alignment, sim_stage1 for maximum face similarity.

It is plug-and-play with the FLUX.1-dev ecosystem, FLUX.1-schnell, ControlNets, LoRAs, and there is an official ComfyUI node plus a Hugging Face demo space. The licensing is split: code is Apache 2.0, but the model weights are CC-BY-NC-4.0 for academic research only. The paper reports beating FLUX.1-dev IP-Adapter and PuLID-FLUX baselines on identity similarity, alignment, and quality, though those are author-run numbers. It sits in the same space as InstantID and PuLID, aimed more at the DiT/FLUX generation of hardware.

Author-reported: surpasses FLUX.1-dev IP-Adapter and PuLID-FLUX on identity similarity, text-image alignment, and quality (Jiang et al., ICCV 2025 Highlight)