SenseNova-U1.5 Full Release: Stronger 4K Editing and Text Rendering

SenseTime has shipped the full release of SenseNova-U1.5-8B-MoT, the production version of the unified multimodal model that was previewed in July. It keeps the same NEO-unify architecture, but the training now builds task-specialized experts (text and infographics, aesthetics, editing) and distills them back into a single model via OPD (on-policy distillation). At inference it is one model with no router.

Six user-visible improvements

The official release focuses on six areas:

• Higher-quality image generation with better composition, color harmony, material rendering, lighting and local detail.
• Better text rendering and infographics, with more legible Chinese and English text in posters, infographics and brand assets.
• More efficient native 4K generation with coherent global structure and stable high-resolution output.
• More reliable native image editing, preserving subject identity and unedited content across local, text, multi-reference, insertion and replacement edits.
• Stronger complex-instruction following for object counts, spatial relationships, layouts and multiple constraints in a single request.
• More precise visual control through bounding boxes, visual markers, and single- or multi-image references.

Benchmarks

Run it in ComfyUI

The official ComfyUI-SenseNova-U1 node pack (v0.3.0) runs the full release locally: text-to-image, single- and multi-image editing with up to 10 ordered reference images, plus accelerated LoRA adapters for local loading. Install it from ComfyUI Manager or the Comfy Registry (comfy node install ComfyUI-SenseNova-U1).

Download SenseNova-U1.5 T2I workflow Download SenseNova-U1.5 editing workflow

Weights: Hugging Face | ModelScope | GitHub