UniGen-AR: Unifying Visual Generation with Auto-Regressive Modeling

Researchers introduced UniGen-AR, pairing a multimodal language model with a next-scale visual auto-regressive decoder.

The single system can handle more than 15 image-valued tasks, including text-to-image generation, editing, and restoration.

It is reported to cut inference latency by up to 19x versus diffusion baselines while matching or improving quality.

Tokenizer design emerged as a key factor for scalability and overall performance.