SenseNova-U1.5: Unified Visual Intelligence Model

What changed

SenseNova-U1.5 represents a significant evolution in native unified multimodal modeling, building upon the encoder-free and VAE-free architecture of its predecessor, SenseNova-U1. The core architectural shift is from independent patch prediction to spatially joint reconstruction. Instead of treating each visual token as an isolated RGB patch, SenseNova-U1.5 projects tokens onto a 2D feature field and uses spatial convolutions and Pixel Shuffle upsampling for progressive image reconstruction. This allows for information exchange between neighboring regions before pixel finalization, leading to improved spatial coherence, color, texture, and geometry resolution, especially at higher resolutions. The model now supports native generation up to 4K resolution.

The training strategy has also been refined from joint reward optimization to a "specialize-then-unify" approach. This involves optimizing dedicated reinforcement learning (RL) experts for distinct capabilities such as visual aesthetics, bilingual text rendering, infographic generation, and image editing. These specialized experts are trained with tailored data, rewards, and sampling strategies. Their consolidated capabilities are then transferred to a single unified policy through multi-expert on-policy distillation, which distills along the student model's own generation trajectories to preserve complementary strengths.

Across extensive evaluations, SenseNova-U1.5 demonstrates advancements in image fidelity, text rendering, complex composition, multi-reference editing, and interleaved generation. It also shows improved instruction following and better preservation of subject identity, geometry, and unmodified regions during editing. Notably, despite limited exposure to structured formats in its generation data, the model effectively generalizes to long, complex, and structured visual instructions, indicating that multimodal understanding can transfer to visual planning and creation.

Why it matters for builders

SenseNova-U1.5 offers builders a more integrated and powerful framework for visual AI tasks. By eliminating the need for separate vision encoders and VAEs, it simplifies the development pipeline and potentially reduces computational overhead. The improved spatial coherence and support for up to 4K resolution open new possibilities for high-fidelity image generation and editing. The "specialize-then-unify" strategy, consolidated through multi-expert distillation, means builders can leverage a single model that excels across a wider range of visual creation tasks, from generating aesthetically pleasing images and accurate bilingual text to creating detailed infographics and performing complex edits.

Practical impact

Builders can explore SenseNova-U1.5's capabilities through its official demo and access its training code, including supervised fine-tuning, reinforcement learning, and on-policy distillation, which will be open-sourced. This release provides an opportunity to experiment with a native unified modeling approach that aims for end-to-end visual perception, reasoning, and creation. Developers can investigate how the spatially joint reconstruction mechanism impacts image quality and coherence compared to previous patch-based methods. The model's generalization to structured visual instructions suggests potential for applications requiring precise layout and composition.

Caveats and source limits

The provided source is a research paper abstract and associated metadata. While it details the architectural changes, training methodology, and claimed performance improvements of SenseNova-U1.5, it does not include specific benchmark results or quantitative comparisons against other models. The exact performance gains in areas like image fidelity, text rendering, and editing are described qualitatively. Information regarding model availability beyond open-sourced training code, such as pre-trained model weights or API access, is not detailed, though HuggingFace collection links are provided. The source also mentions a "limited exposure to structured formats" in training data, which is presented as a challenge overcome by the model's generalization, but the extent of this limitation and its potential impact on specific edge cases are not fully elaborated.