商汤发布SenseNova U1 Pro：原生8K多模态模型，自我修正数十轮

SenseTime has released SenseNova U1 Pro, a multimodal model that renders images at 8K natively andself-corrects across dozens of reasoning rounds.

Every image generator you have used so far is blind after it starts, it never looks back.

So you are the one who spots the mistake and hits regenerate. U1 Pro moves that checking step inside the model, which means it grades its own picture and redraws until it matches your brief.

• One model, not two: NEO-Unify puts understanding, generation, and action in the same network, so the part that reads your brief is the part that draws it, instead of handing a compressed instruction to a separate pixel generator that never saw the original request.
• The Agentic Generation Loop: it can look at what it just produced, compare that against what you asked for, and redraw across dozens of rounds, which is the mechanism behind the low error rate claimed on rendered text.
• Native 8K, including ultra-wide and ultra-tall shapes, so a large format image leaves the model at final size rather than passing through a separate upscaler that invents detail the model never reasoned about.