Learning visual concepts before rendering improves image generation in controlled tests

Mao and colleagues combine autoregressive pretraining with diffusion rendering, reporting strong benchmark scores from a roughly three-billion-parameter model trained with 6,600 H100 GPU-hours.

Big Tech

Yucheng Mao · Chongjian Ge · Zhuowen Tu · Hao Tan · Hanwen Jiang

University of California, San Diego · Adobe Research

Research Digest·28 October 2026·2 min read

Illustration: Learning visual concepts before rendering improves image generation in controlled tests

In one line

Visual generation is better when autoregressive pretraining learns semantic tokens and diffusion post-training renders them.

Before you read this

Three ideas this paper builds on

role decouplingSeparating distinct functions such as planning, reasoning, and execution into separate components improves efficiency, quality, and reliability.38 papers →diffusion modelA type of generative model that learns to reverse a gradual noising process to create data.151 papers →scaling lawA predictable power-law relationship between compute or data and model performance.16 papers →

Read it as

To make an image from text, a model must understand objects and place every pixel. Most models learn both with one method, which is expensive. This work splits those jobs into a concept stage, which predicts compact symbols of image content, and a rendering stage, which cleans noise into a picture. The concept stage trains only on text descriptions, not by rebuilding images.

Controlled tests showed the concept stage should maximize meaning, not pixel reconstruction. Scaling rules say how much compute goes to each stage.

The final model scored 86 on GenEval and 87 on DPG-Bench, standard text-to-image tests, using far less training compute than a comparable model. So know what to draw before drawing it.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

role decoupling autoregressive pretraining diffusion model semantic tokenizer scaling law flow matching