Paper Explained: SenseNova-U1.5 — One Model That Both Sees and Draws, With No Encoder and No VAE

SenseNova-U1.5 is an 8B unified model that learns straight from pixels and words, with no vision encoder for understanding and no VAE for generation. We walk through its seam-free spatial decoder and its 'train specialists separately, then distill them into one' post-training, following the paper itself.

SenseNova-U1.5: Towards Native Unified Visual Intelligence

Merging the "viewer" and the "illustrator" into one person

The paper is "SenseNova-U1.5: Towards Native Unified Visual Intelligence" (arXiv:2609.11929, published September 10, 2026).

Here is what the abstract says. SenseNova-U1.5 is an 8B-MoT (Mixture-of-Transformers) native unified multimodal model that understands, reasons about, and generates visual content, built without any vision encoder or VAE. The authors strengthen its visual interface with spatially coherent patch reconstruction, and scale training with curated generation and editing data, better task formulation, structured prompt enhancement, and native resolutions up to 4K. In post-training they optimize separate experts for aesthetics, bilingual (Chinese/English) text rendering, infographic generation, and image editing, then consolidate them through multi-expert on-policy distillation. They report large gains in image fidelity, text rendering, complex composition, multi-reference editing, and interleaved text-image generation, along with better instruction following and better preservation of subject identity, geometry, and unedited regions. Although structured formats are rare in its generation data, the model generalizes to long, complex, structured visual instructions, which the authors read as evidence that multimodal understanding can transfer to visual planning and creation. Training code for SFT, RL, and on-policy distillation is announced for release.

Most image AI systems today employ two translators. One turns images into meaning (a vision encoder); the other turns meaning back into images (a VAE). They work from different dictionaries, so every time the model draws what it understood, something has to be looked up again. That is the paper's starting point (§1): understanding lives in a space tuned for semantic abstraction, generation in a space tuned for pixel fidelity.

SenseNova-U1.5 is the latest in a line of models that drop the translators and hire one person who reads and writes pixels and words directly. Its predecessor, SenseNova-U1, showed the approach was feasible. U1.5 fixes two of its weaknesses: visible seams between patches, and the way different skills interfere with each other when trained under a single policy.

Intuition: a 32×32 block is one word

U1.5 cuts the image into $32\times 32$-pixel blocks and treats each block as a single token, one "word" from the model's point of view (§3.1). Two convolutions shrink the image by 16× and then 2× to get there. A 1024×1024 image becomes 32×32 = 1,024 tokens. Because there is no encoder in the way, the authors argue, the image becomes a short sequence while losing almost nothing.

The trouble was at the exit. U1 used a small MLP to turn each token back into a 32×32 RGB patch on its own. A patch decided its colors without knowing what its neighbor would do, so colors and textures drifted apart at the borders. The paper describes this as seams, grid artifacts, and texture breaks that get worse at high resolution (§1, §3.1).

JPEG has the same kind of problem. It splits the image into 8×8 blocks and processes each one independently, so heavy compression makes the block edges show. Lower the quality in the figure below and watch them appear.

U1.5's fix is to lay the tokens back out as a 2D map and let neighbors exchange information before any pixel is finalized.