ByteDance recently launched an image generation tool called InfiniteYou (InfU), the core function of this tool is to transform text descriptions into high-quality images with personal identity characteristics. Unlike traditional face-changing applications, InfiniteYou can not only flexibly change scenes and content, but also accurately retain the user's identity characteristics, so that the generated images are both consistent with the description and highly personalized.
The core of InfiniteYou is InfuseNet, a technology that can inject user identity characteristics into the advanced image generation model of Diffusion Transformer (DiT). Through the fine operation of "residual connection", InfuseNet not destroys the original generation ability while enhancing the face similarity, thereby achieving high-quality image generation.
In addition, InfiniteYou's R&D process went through multi-stage training, including pre-training and supervised fine-tuning (SFT) using synthetic single-person multi-sample (SPMS) data. This refined training strategy significantly improves the alignment of text and images, making the generated images more in line with text description, and at the same time improves image quality and aesthetics, effectively alleviating the common facial copy-paste problems after "face change".
ByteDance has also released two model versions: aes_stage2 and sim_stage1, focusing on text and picture alignment and face similarity respectively. Users can choose the appropriate version according to their needs. Through comparative experiments, InfiniteYou surpasses existing advanced methods in terms of identity similarity, text image alignment, image quality and aesthetics, such as FLUX.1-dev IP-Adapter and PuLID-FLUX.
InfiniteYou also has the "plug and play" feature, which can seamlessly integrate with various FLUX.1-dev variants, ControlNets and LoRAs and other existing tools, providing stronger controllability and customization capabilities. This strong compatibility will make a valuable contribution to the wider community.
It should be noted that InfiniteYou is currently released based on Creative Commons Attribution-NonCommercial4.0 International Public License and is for academic research only. Downloading and using related models (such as InsightFace's face model, FLUX.1-dev base model, LoRA, etc.) must comply with its original permission. At the same time, developers also hope that users can abide by local laws and regulations, use this technology responsibly, and avoid any potential abuse.
Project entrance: https://top.aibase.com/tool/infiniteyou