SenseTime open-sources 8B unified multimodal SenseNova U1.5 Lite with native 4K output
On Aug 21 SenseTime released the 8B unified multimodal SenseNova U1.5 Lite under open source, native support for 3–4K instruction length and native 4K output, focusing on complex instructions, visual quality, text/layout, and native image editing. Code is on GitHub, Hugging Face, and ModelScope.
On August 21, SenseTime announced the formal-version open-source release of SenseNova U1.5 Lite, a lightweight unified multimodal large model. Built on the in-house NEO-unify architecture, the model uses MOPD (Multi-Expert Online Policy Distillation) to fuse multiple expert capabilities into a single 8B model that runs smoothly on a single consumer GPU, avoiding the system overhead and inference latency of explicit expert routers.
Compared with the prior preview, the formal version retrains on data and post-training tailored to real visual tasks and reinforces five capabilities: native support for 3–4K character complex instructions with simultaneous subject, quantity, spatial-relationship, text, layout, and style constraints; higher visual generation quality with improved composition, color, material, lighting, realism, and local detail; more reliable native image editing that preserves subject identity, spatial structure, layout relations, and non-edited regions; stronger text and complex layout (Chinese/English copy, posters, infographics, brand visuals, multi-text composition); fine-grained visual control via Bounding Box, Visual Marker, and single/multi-image references; native 4K high-resolution output that balances overall composition with micro text, material texture, and light refraction.
In real-world scenarios the model demonstrates stability and controllability on creative posters, dense infographics, artistic-feel generation, multi-reference merging, local edits, and text replacement. The model is open-sourced globally and available on GitHub (github.com/OpenSenseNova/SenseNova-U1), Hugging Face (huggingface.co/collections/sensenova/sensenova-u15), and ModelScope (modelscope.cn/models/SenseNova/SenseNova-U1.5-8B-MoT), with online playground access in SenseNova Studio. SenseTime states the model surpasses class peers in instruction following and image-edit preservation, and rivals super-scale commercial models in text rendering and complex layout delivery.