UniTok

Why we included this project

UniTok is a single visual tokenizer built to feed both generation and understanding models, so LlamaGen and LLaVA can draw on the same discrete tokens instead of each needing its own encoder. It came out of a NeurIPS 2025 Spotlight paper and backs the Liquid framework, where the authors built a unified multimodal model that they report sets a new state of the art among autoregressive MLLMs. One finding stands out for practical use: UniTok generates well without classifier-free guidance, cutting gFID on ImageNet 256x256 from 14.6 to 2.51 when paired with a LlamaGen-XXL generator. Model checkpoints, a Hugging Face demo, and evaluation configs are available if you want to verify that yourself.