SenseTime Releases U1Pro Multimodal Creation Model

新智元 · wechat · 2026-07-18

At WAIC, SenseTime released its next-generation multimodal model, SenseNova U1Pro, featuring a unified base for "understanding, generation, and action." The article primarily showcases its image generation capabilities in tasks like 8K scrolls, movie posters, infographics, Chinese long-text typography, and character design sheets, emphasizing that it acts more like a "thinking designer" rather than a slot-machine-style image generator.

Technically, U1Pro employs a proprietary unified architecture called NEO-unify, integrating understanding, generation, and action into the same representation and Transformer sequence. The creation process is completed through long-range iterations similar to an Agentic Generation Loop. The article also notes:

Training incorporates a "Chain-of-Thought instruction" and multi-dimensional rewards to enhance composition, text accuracy, design aesthetics, and consistency
By utilizing larger patches and hierarchical noise control, it reduces the context and compute pressure of direct 8K output
The company invited 200 art academy students and designers for review, claiming that currently, only U1Pro and GPT-Image-2 meet the threshold for being "directly deliverable to clients"
Future goals extend beyond visual creation to spatial intelligence, embodied brains, urban planning, and architectural design

Original post →

SenseTime Releases U1Pro Multimodal Creation Model

More from Models