Researchers at ByteDance Clever Creation have developed a brand new AI mannequin that generates a number of variations of an identification together with its paper, demo, and code. Liming Jiang, a senior analysis scientist at ByteDance, made the announcement on X on Sunday.
The brand new AI mannequin referred to as InfiniteYou (InfU) goals to handle the challenges of identity-preserved picture era. One can create a number of variations of their identification in several settings through the use of prompts as required, guaranteeing good accuracy. The mannequin leverages Diffusion Transformers (DiTs) to generate pictures that not solely keep the identification of an individual from a supply {photograph} but in addition enable for versatile text-based modifying.
InfU goals to beat the constraints present in present strategies, comparable to inadequate identification similarity, poor text-image alignment, and low era high quality. The core of InfU is InfuseNet, a part designed to inject identification options into the DiT base mannequin by means of residual connections. This course of enhances identification similarity whereas preserving the mannequin’s generative capabilities.
To additional refine the mannequin’s efficiency, a multi-stage coaching technique was employed, incorporating pretraining and supervised fine-tuning (SFT) with artificial single-person-multiple-sample (SPMS) knowledge. The coaching method was designed to enhance text-image alignment, improve picture high quality, and mitigate face copy-pasting points.
The official web site talked about, “InfU contains a fascinating plug-and-play design appropriate with many present strategies. It naturally helps base mannequin substitute with any variants of FLUX.1-dev, comparable to FLUX.1-schnell for extra environment friendly era.”
“The compatibility with ControlNets and LoRAs supplies extra controllability and suppleness for customised duties. Notably, the compatibility with OminiControl extends our potential for multi-concept personalisation, comparable to interacted identification (ID) and object personalised era,” the paper added.
The code is obtainable on the GitHub web page, and one can entry the demo and the mannequin on Hugging Face to strive it out.
ByteDance has been making a number of developments in 2025, together with Goku as a substitute for Google’s Luma and a React Native killer. The AI mannequin provides to its checklist of thrilling developments up to now.
