AIBRIEF
SenseTime's new 8 billion parameter image model, SenseNova U1.5 Preview, ships open source, but the model's own documentation flags it a preview. Here's what it does and where the demos stop.
A hand-drawn recipe card rebuilt from a photo, then redrawn region by region with a red box and a sentence, that's what SenseTime's new open-source model does at native 4K pixels.
SenseTime released SenseNova U1.5-Lite-Preview, an 8B-parameter multimodal model that unifies generation and editing in a single network. Weights are on Hugging Face; docs sit in the GitHub repo, both under an open license.
The release covers visual understanding, reasoning, generation, and editing. Editing primitives include reference images, multi-image composition, local text edit, and red-box or coordinate-precise edits. A QbitAI walkthrough demos a postal-mail flowchart, a museum exhibit remade in a new design language, a Chinese character repainted stroke by stroke, and a resume generated from a portrait.
The 8B size reaching native 4K is an aggressive claim. The deltas 163.com reports (Qwen-Image-Bench Q 47.14 to 55.20; GEdit-Bench EN 7.47 to 8.17) come straight from the company release. The model card and GitHub docs explicitly label U1.5 a preview ahead of a more capable production version, and community feedback on the prior U1 drop flagged photorealism at SD15/SDXL level: fast on text and infographics, weaker on portraits.
Real capability release. Every demo vendor-curated. Production version still ahead.