TL;DR One auto-regressive model handles 15+ visual generation tasks — text-to-image, editing, restoration, and perception — at up to 5× lower latency and >10× fewer FLOPs than diffusion, trained only on public data.
Modern vision pipelines stay fragmented — generation, editing, restoration, and perception each handled by separate models. We study Unified Visual Generation (UVG), where one model produces diverse image-valued outputs through a single multimodal interface.
UniGen-AR pairs a general-purpose multimodal language model encoder with an efficient next-scale visual auto-regressive (VAR) decoder. The MLLM turns free-form instructions and reference images into a unified conditioning sequence that guides the VAR decoder across 15+ tasks in four families.
The result: up to 5× lower inference latency than diffusion baselines while matching or improving quality. Our analysis identifies VQ-VAE tokenizer design — codebook size and hierarchy — as the critical factor for scaling VAR in UVG.
A Qwen2.5-VL encoder reads the instruction and reference image into text embeddings; a learnable [SoS] token and cross-attention steer an Infinity-style VAR decoder that predicts discrete visual tokens, decoded to pixels by a VQ-VAE.
Under matched conditioning and hardware, swapping a diffusion decoder for the VAR decoder cuts inference latency dramatically — while improving GenEval quality.
Qwen-2.5-VL conditioning · 512×512 · single H100
Browse outputs by family. Editing and restoration show input → UniGen-AR output.
Input · UniGen-AR (Ours) · task-specialized and unified baselines, across perception, restoration, and editing.
@article{unigenar2026,
title = {UniGen-AR: Unifying Visual Generation with Auto-Regressive Modeling},
author = {Bao, Zhipeng and Zhu, Zhen and Kumari, Nupur and Bagchi, Anurag
and Wang, Yu-Xiong and Tokmakov, Pavel and Hebert, Martial},
journal = {arXiv preprint arXiv:2607.24157},
year = {2026}
}