SenseNova-U1.5: Towards Native Unified Visual Intelligence
Haiwen Diao, Jiahao Wang, Chenjing Ding, Hanming Deng, Jiangnan Chen, Ruixi Zhang, Ruohui Wang, Wenwen Tong and 57 others
Published September 10, 2026 · on the feed September 11, 2026 · ▲ 273 · arXiv 2609.11929
The abstract
We launch SenseNova-U1.5, an 8B-MoT native unified multimodal model that understands, reasons about, and generates visual content within an encoder-free and VAE-free architecture. We strengthen its visual interface through spatially coherent patch reconstruction and scale its training with carefully curated generation and editing data, improved task formulation, structural prompt enhancement, and native resolutions of up to 4K. For post-training, we optimize specialized experts for visual aesthetics, bilingual text rendering, infographic generation, and image editing, and consolidate their capabilities through multi-expert on-policy distillation. Across extensive evaluations, SenseNova-U1.5 largely advances image fidelity, text rendering, complex composition, multi-reference editing, and interleaved generation, while improving instruction following and preserving subject identity, geometry, and unmodified regions. Despite limited exposure to structured formats in its generation data, SenseNova-U1.5 generalizes effectively to long, complex, and structured visual instructions, further proving that multimodal understanding can transfer to visual planning and creation. Together, these findings position native unified modelling as a promising path towards systems that perceive, reason and create within a fully end-to-end framework. We will open-source training code, including supervised fine-tuning, reinforcement learning, and on-policy distillation.
The Brief
2:39A short podcast summary of the paper. Written and voiced by AI.
Transcript
SenseNova-U1.5. Towards Native Unified Visual Intelligence. By Haiwen Diao and sixty-four others. Published on the tenth of September. Its code repository is worked out: about three hundred source files, with a readme and a licence.
A model that understands images, reasons about them, and also generates them usually does so with separate machinery for seeing and for drawing. SenseNova-U1.5 aims for a native unified model instead: one that does all three within an architecture free of both a separate encoder and a VAE. Unifying them means one set of weights has to do perception, reasoning and creation at once, and has to take in images and put them out without separate components on either side.
It is an eight billion parameter MoT model, understanding, reasoning about, and generating visual content in one end-to-end system. Its visual interface is strengthened through spatially coherent patch reconstruction. MoT describes the model's architecture, and the claim is that a model of this size can cover the whole range from looking at an image to making one.
Training is scaled with carefully curated data for generation and editing, an improved formulation of tasks, structural enhancement of prompts, and native resolutions of up to 4K. For post-training, specialized experts are optimized separately for visual aesthetics, bilingual text rendering, infographic generation, and image editing. Their capabilities are then consolidated into one model through multi-expert on-policy distillation. The pattern here is to build skill in parts and then merge it: separate experts get good at one thing each, and distillation folds what they learned back into a single model.
Across extensive evaluations, SenseNova-U1.5 substantially advances image fidelity, text rendering, complex composition, editing with multiple references, and interleaved generation. It also improves instruction following, while preserving subject identity, geometry, and regions that were not meant to change.
One result stands out. Its generation data had limited exposure to structured formats, and yet the model generalizes effectively to long, complex, structured visual instructions. The authors take this as evidence that multimodal understanding can transfer to visual planning and creation. Put differently, the model was not taught much about following structured layouts when it was learning to draw, but what it understands about such layouts carried over to drawing them.
The authors position native unified modelling as a promising path toward systems that perceive, reason and create within one fully end-to-end framework. Training code, covering supervised fine-tuning, reinforcement learning and on-policy distillation, is to be open-sourced.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
One model to see and to draw. The trouble starts at the seams.
You want one model that understands images, reasons about them and draws them. Most systems see through one component and draw through another, so seeing and drawing live in different spaces.
The road ahead
- 1 · The problem ~7 min One model should see and draw images, but common designs split seeing from drawing, and a single-model attempt drew each patch alone, leaving seams.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Rebuild pixels jointly on a 2D grid so neighbours agree, train a specialist per skill, then distil all specialists into one student on its own outputs.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Rebuild pixels jointly, scale training on curated data, train one expert per skill, then distil the experts into one model on its own outputs.
- 6 · The experiment ~6 min You’ll guess: SenseNova-U1.5 saw few structured formats in its generation data. How did it cope with long, complex, structured visual instructions?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.