SenseNova-U1.5 unifies image understanding and generation without encoders or VAEs

A paper introduces SenseNova-U1.5, described as an 8B-MoT native unified multimodal model, meaning a single system that understands, reasons about, and generates visual content, built on what the authors call an encoder-free and VAE-free architecture. Most systems that both interpret and produce images stitch together a separate vision encoder for understanding and a separate generative decoder, often built on a VAE, for output; SenseNova-U1.5's authors say they instead strengthen the model's visual interface directly, through what they call spatially coherent patch reconstruction. They scale training with curated generation and editing data, an improved task formulation, structural prompt enhancement, and native training resolutions of up to 4K, rather than upscaling after the fact.

For post-training, the authors optimize separate specialized experts, one each for visual aesthetics, bilingual text rendering, infographic generation, and image editing, then fold those experts' capabilities back into a single model through a process they call multi-expert on-policy distillation.

Across what the paper calls extensive evaluations, the authors report that SenseNova-U1.5 largely advances image fidelity, text rendering, complex composition, multi-reference editing, and interleaved generation, while also improving instruction following and better preserving subject identity, geometry, and regions of an image that an edit should leave unmodified. The text does not name any benchmark, comparison baseline, or numeric score behind these claims; the improvements are described only in these qualitative terms.

The authors also report that despite limited exposure to structured formats in its generation training data, SenseNova-U1.5 generalizes effectively to long, complex, and structured visual instructions. They treat that result as further evidence that multimodal understanding can transfer into visual planning and creation, and on that basis frame native unified modelling as a promising, though not yet established or definitive, path toward systems that perceive, reason, and create within a single end-to-end framework.

On availability, the team states only that it will open-source its training code, covering supervised fine-tuning, reinforcement learning, and on-policy distillation. That is a stated future commitment rather than something already released: the text gives no date and no repository or release channel, and it does not say whether SenseNova-U1.5 itself, meaning the model or its weights, will be made public at all.

Key facts

SenseNova-U1.5 is an 8B-MoT native unified multimodal model that understands, reasons about, and generates visual content in a single encoder-free, VAE-free architecture, with training scaled to native resolutions of up to 4K.
Post-training uses specialized experts for visual aesthetics, bilingual text rendering, infographic generation, and image editing, merged into one model through what the authors call multi-expert on-policy distillation.
The authors report that across extensive evaluations, the model largely advances image fidelity, text rendering, complex composition, multi-reference editing, and interleaved generation, without naming any benchmark or numeric score.
Despite limited exposure to structured formats during training, the authors say the model generalizes to long, complex, structured visual instructions, which they read as evidence that multimodal understanding transfers to visual planning and creation.
The team says it will open-source the training code, including supervised fine-tuning, reinforcement learning, and on-policy distillation, but gives no release date or repository, and does not state whether the model itself will be made public.

Why it matters

Systems that both interpret and generate images typically bolt together two separate components: a vision encoder to understand pictures and a generative decoder, often a VAE, to produce them. SenseNova-U1.5's authors describe an architecture that drops both pieces, folding understanding, reasoning, and generation of visual content into one encoder-free, VAE-free model instead. If that holds up, it points toward a simpler stack for building systems that perceive and create images together rather than pairing separate models for each half of the job. The paper also targets specific weak spots in image generation: getting text legible inside a generated image in two languages, producing usable infographics, and making multi-reference edits that leave a subject's identity, geometry, and untouched regions intact, while training natively at resolutions up to 4K instead of upscaling afterward.

Who it affects

This is a research paper describing an architecture and a training recipe, not a released product, so the immediate audience is researchers and engineers building unified multimodal systems, ones that need a single model to both interpret and generate visual content rather than chaining separate tools together. The specific capabilities the authors describe, bilingual text rendering, infographic generation, and multi-reference editing that preserves a subject's identity and geometry, point to teams whose downstream use cases involve generating images with legible embedded text or making complex, targeted edits, though nothing here is yet packaged for a general user to try.

How to use it

There is nothing to install today. The text describes the architecture, training approach, and evaluation results, but the only concrete availability commitment is that the team will open-source its training code, covering supervised fine-tuning, reinforcement learning, and on-policy distillation. No date is given for that release, and no repository or release channel is named. The text also does not say whether SenseNova-U1.5 itself, the model or its weights, will be released publicly at all, only that the training code will be; anyone wanting to use the model as described has nothing yet to run.

How solid is it

The claims here are the authors' own, drawn from what the text calls extensive evaluations, and reported only qualitatively: the model 'largely advances' several capabilities and 'improves' others, with no benchmark named, no baseline system identified, and no numeric score given for any of it. No individual authors or institution are named in the text, which is written entirely in the first person plural ('we'), so there is no byline or affiliation to check the claims against. The architecture tag itself, '8B-MoT', is stated without saying what unit '8B' counts or what 'MoT' stands for, leaving even the model's basic size description less precise than it looks at first glance.

Risks and caveats

Every improvement claim in the paper, image fidelity, text rendering, composition, editing, generalization to structured instructions, rests on the authors' own qualitative description of their own evaluations, with no named benchmark, baseline, or score to weigh it against independently. The claim that the model generalizes to long, complex, structured visual instructions despite limited exposure to such formats in training is the authors' own interpretation of a single training run, not a result checked against an alternative setup. And since only the training code has a stated release commitment, with no date or repository given, and the model's own public availability is not addressed at all, none of these claims can yet be tested independently.

“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.”

— the paper