SenseNova-U1.5: Towards Native Unified Visual Intelligence
Haiwen Diao, Jiahao Wang, Chenjing Ding, Hanming Deng, Jiangnan Chen, Ruixi Zhang, Ruohui Wang, Wenwen Tong, Xiangyu Fan, Yubo Wang, Yue Zhu, Yuwei Niu, Zhengqi Bai, Zhiqian Lin, Zhitao Yang, Zhongang Cai, Bo Yang, Chen Feng, Chengguang Lv, Guangjia Liu, Guanlin Wang, Hanyu Zhang, Haojia Yu, Hongcan Xiao, Hongli Wang, Huan Wu, Huaping Zhong, Jian Fang, Jianan Fan, Jiaqi Li, Jiefan Lu, Jing Zuo, Jingcheng Ni, Junxiang Xu, Linjun Dai, Mutian Xu, Peishen Yan, Penghao Wu, Ruijie Mao, Ruisi Wang, Shihao Bai, Shuang Yang, Shuya Yang, Shuyan Zheng, Silei Wu, Siying Li, Tao Chu, Tianbo Zhong, Tongxi Zhou, Weichao Luo, Weichen Fan, Wenhao Jia, Wenjie Gao, Xiangli Kong, Yan Li, Yang Yong, Zimo Wen, Zixuan Qian, Wenxiu Sun, Ruihao Gong
SenseNova-U1.5 is an 8B-parameter native multimodal model that unifies visual understanding and generation without encoders or VAEs.
How can we unify visual understanding and generation into a single, encoder-free, VAE-free architecture using a Mixture-of-Transformers (MoT) approach?
Most multimodal systems rely on separate pathways for perception and generation, forcing a constant, lossy conversion between encoder features and latent spaces. This architectural divide prevents seamless coordination between visual reasoning and pixel-level synthesis. SenseNova-U1.5 replaces these parallel pathways with a single, end-to-end Transformer backbone that learns directly from pixels and text. It uses a spatially joint decoder to resolve local image continuity and a multi-expert distillation strategy to consolidate specialized generative capabilities. Across standard benchmarks, this unified approach achieves state-of-the-art performance among open-source models in image fidelity, bilingual text rendering, and complex compositional instruction following.
Paper Primer
The model hinges on two core shifts: moving from independent patch prediction to spatially joint reconstruction, and replacing joint reward optimization with a "specialize-then-unify" strategy. The decoder is like a mosaic artist who, instead of placing tiles in isolation, uses a grid-based framework to ensure colors and textures flow smoothly across neighboring boundaries.
SenseNova-U1.5 achieves superior compositional consistency and instruction following compared to its predecessor and larger open-source baselines.
Quantitative results on GenEval and GenEval2 benchmarks, where the model demonstrates improved attribute binding, counting, and verb-related generation. It reaches an overall score of 0.92 on GenEval, outperforming both SenseNova-U1 and larger models like Qwen-Image.
The model maintains strong language and multimodal reasoning capabilities despite the addition of high-fidelity generation.
Performance on MMLU-Pro (86.67) and C-Eval (90.41), alongside competitive results on standard VQA and reasoning benchmarks. Instruction-following performance on IFEval improved to 93.35, showing no dilution of reasoning competence.
Why use a "specialize-then-unify" strategy instead of training one model on all tasks at once?
Jointly optimizing competing objectives—such as aesthetic synthesis and infographic design—can entangle goals and dilute task-specific gains. Specializing experts first allows for tailored rewards and sampling, which are then consolidated into a single policy via on-policy distillation.
What is the primary advantage of the "native" architecture over traditional modular systems?
It eliminates the need for external vision encoders and variational autoencoders, allowing the model to process visual inputs and generate pixels within a single representational space without repeated feature conversion.
Unified Visual Intelligence
SenseNova‑U1.5 unifies visual understanding, reasoning, and generation in a single encoder‑free, VAE‑free model.
SenseNova‑U1.5 replaces fragmented multimodal pipelines with a single Mixture‑of‑Transformers substrate that handles visual understanding, reasoning, and generation without separate encoders or VAEs.
It treats visual perception, reasoning, and creation as a single computational task, eliminating the need for separate encoder, decoder, or compression stages.
**Figure.** Performance comparison across various benchmarks using radial bar charts.
**Figure 1.** Showcases of SenseNova-U1.5 in infographics and human generation.
The shift from encoder‑based pipelines to a native unified architecture is the key driver of the reported gains.
Architecture and Training
Core mechanisms that unify visual understanding and generation in a single native model.
Fragmented pipelines that route images through pretrained encoders and then back through VAEs force perception and synthesis into disjoint representational spaces, limiting seamless coordination of understanding, reasoning, and creation.
MoT treats each visual or linguistic token as an expert that can specialize on a subset of tasks, while a shared backbone routes information among them, enabling a single model to handle both understanding and generation.
How does MoT differ from a conventional Mixture‑of‑Experts that uses a learned gating network?
Conventional MoE routes every token through a soft‑max gate that selects a subset of experts based on the token’s content, which introduces stochastic load‑balancing and extra communication. MoT uses a deterministic token‑type mask: clean image tokens always use the “understanding” expert, noisy generation tokens always use the “generation” expert. This eliminates gating overhead and guarantees that the right expert processes each modality.
Instead of compressing an image into a latent vector, the interface slices the raw picture into 32 × 32 patches and encodes each patch directly into a single token, preserving almost all pixel information while keeping the sequence short.
Input image (64 × 64 × 3) → Conv stride 16 → 4 × 4 × C feature map.
Second Conv stride 2 expands the map to 8 × 8 × C, then a linear projection collapses each 32 × 32 region into a single token.
Four tokens are produced, each carrying the raw visual content of its patch plus a 2‑D positional code.
The token sequence length is 4, far smaller than the 4 096 raw pixels.
This tokenization preserves pixel‑level detail (near‑lossless) while reducing the sequence length enough for large‑scale multimodal training.
Why does encoding each 32 × 32 patch as a single token avoid the information loss typical of VAEs?
VAEs compress an entire image into a low‑dimensional latent, discarding high‑frequency details. Here the patch‑wise linear projection retains the raw pixel values (up to the convolutional receptive field) and only reduces dimensionality via a learned linear map, so the reconstruction can recover the original colors and textures with minimal degradation.
When denoising images of different sizes, the amount of injected noise must scale with resolution; the model therefore receives an explicit embedding that tells it how much noise to expect for a given (H, W).
Compute $\bar{\sigma}_R = 0.8 / 1.0 = 0.8$ for the 2048 × 2048 case.
Pass 0.8 through NSEmb → embedding vector $e$₁.
Add e₁ to the timestep embedding $\tau_{t}$ to form $s_t$ for the denoiser.
For 4096 × 4096, $\bar{\sigma}_R = 1.0$ → embedding e₂, which is slightly larger, signalling more noise.
The embedding scales the denoiser’s expectations proportionally to image size, preventing under‑denoising at high resolutions.
What would happen if the model ignored resolution when conditioning the noise scale?
At 4K resolution the injected noise would be too low relative to the pixel count, so the denoiser would receive an overly clean signal and fail to learn the correct diffusion dynamics, leading to blurry or artifact‑prone generations.
After the backbone produces a 2‑D feature map, the decoder progressively upsamples it with Pixel‑Shuffle layers interleaved with 3 × 3 convolutions, allowing neighboring patches to share information before the final RGB pixels are emitted.
First Pixel‑Shuffle (×2) converts (1, 64, 8, 8) → (1, 16, 16, 16) by moving 4 channels into each spatial dimension.
3 × 3 convolution on (1, 16, 16, 16) mixes neighboring patches.
Second Pixel‑Shuffle (×2) yields (1, 4, 32, 32).
Another 3 × 3 convolution smooths the 32 × 32 feature map.
Final Pixel‑Shuffle (×8) expands to (1, 1, 256, 256), which is then projected to RGB.
The staged upsampling lets the model refine local details at each resolution step, preventing the abrupt seams that arise when each patch is decoded independently.
Why combine Pixel‑Shuffle with convolutions instead of using a single upsampling layer?
Pixel‑Shuffle alone merely rearranges channels, preserving the independence of each patch. The interleaved convolutions allow information to flow across patch boundaries at intermediate resolutions, which is essential for coherent textures and smooth edges in high‑resolution outputs.
Four task‑specific RL experts (Aesthetic, OCR, Infographic, Editing) are first trained on their own data and rewards; their policies are later merged into a single model via on‑policy distillation, preserving each specialty while yielding a unified capability.
Why not train a single RL policy on all tasks simultaneously?
Joint RL would force a single reward signal to accommodate conflicting objectives (e.g., aesthetic quality vs. precise OCR), leading to sub‑optimal performance on each task. By specializing first, each expert can fully exploit its own reward landscape; the subsequent distillation then merges the strengths without the interference that a monolithic RL would suffer.
The training pipeline proceeds through five stages: (1) generation pre‑training with flow‑matching and perceptual loss, (2) unified mid‑training that mixes understanding and generation data, (3) supervised fine‑tuning on instruction data, (4) task‑specific RL for four experts, and (5) on‑policy distillation that consolidates the experts into the final unified model.
**Figure 2.** Showcases of SenseNova-U1.5 in image editing and multi-reference generation.
**Figure 3. Overview of SenseNova-U1.5.** Compared with SenseNova-U1, U1.5 further refines the near-lossless visual interface on both encoding and decoding: resolution-aware noise conditioning is extended to the 4096 × 4096 range, while the original patch-wise MLP head is replaced by a lightweight spatial decoder with Pixel Shuffle and 3 × 3 convolutions. These improvements preserve compact 32 × 32 visual tokenization while enhancing spatial continuity, high-resolution fidelity, and downstream robustness.
Table 1 lists the concrete configuration of SenseNova‑U1.5 (32 × 32 patch size, 42 layers, 32 / 8 heads, 8.2 B parameters), while Table 2 details the staged training recipe, including learning‑rate schedules, sequence lengths, and frozen modules per stage.
**Figure 4.** Post-training process of SenseNova-U1.5, including multi-expert reinforcement learning and on-policy distillation.
**Figure 5.** Training corpus of SenseNova-U1.5. From left to right, the charts show the hierarchical composition of the Image Generation, Image Editing, Interleaved, and RL Training datasets. The inner ring denotes major data categories and their proportions, while the outer ring further decomposes them into fine-grained subclasses. Together, these distributions highlight the broad coverage of natural and synthetic imagery, diverse editing scenarios, interleaved image-text and video content, and post-training data.
SenseNova‑U1.5’s core trick—spatially joint reconstruction via a Pixel‑Shuffle decoder combined with a token‑type‑masked MoT—delivers near‑lossless visual fidelity while unifying perception and generation in a single native model.
Multimodal Understanding Performance
Evaluates SenseNova‑U1.5’s multimodal understanding and image generation against strong baselines.
This section reports how SenseNova‑U1.5 performs on multimodal understanding and image‑generation benchmarks, directly comparing against prior models.
SenseNova‑U1.5 maintains multimodal performance with no degradation despite added generation capabilities.
Table 3 shows strong scores across STEM reasoning, VQA, OCR, hallucination, and visual reasoning, matching or exceeding prior baselines.
**Table 3.** Quantitative evaluation results on multimodal and language understanding benchmarks. Note that * denotes results reproduced using VLMEvalKit [33]. SenseNova-U1, SenseNova-U1.5, and Qwen3-VL are all initialized from Qwen3 LLM weights [143]. Besides, Gemma4-12B [113] represents a recent state-of-the-art encoder-free model for visual understanding.
**Table 4.** Quantitative evaluation results on Qwen-Image-Bench-EN. Results are reported on the English subset. # Params denotes the number of parameters in the generation component; A in this column denotes activated parameters during inference.
**Table 5.** Quantitative evaluation results on Qwen-Image-Bench-ZH. Results are reported on the Chinese subset. # Params denotes the number of parameters in the generation component; A in this column denotes activated parameters during inference.
Generation Benchmarks
Benchmark results show SenseNova‑U1.5 leads open‑source models across alignment and generation metrics.
SenseNova‑U1.5 replaces fragmented encoders with a unified Mixture‑of‑Transformers that handles visual understanding and generation jointly. The following results evaluate how this design translates into benchmark performance.
SenseNova‑U1.5 leads open‑source models on GenEval with an overall score of 0.92.
Table 6
**Table 6.** Quantitative evaluation results on GenEval. The parameters of the generation component are denoted as # Params.
**Table 7.** Quantitative evaluation results on GenEval2. The parameters of the generation component are denoted as # Params.
**Table 8.** Quantitative evaluation results on OneIG-EN. The parameters of the generation component are denoted as # P.
**Table 9.** Quantitative evaluation results on OneIG-ZH. The parameters of the generation component are denoted as # P.
**Table 10.** Quantitative evaluation results on DPG-Bench. The parameters of the generation component are denoted as # Params
Qwen3‑VL extends the Qwen language model with a visual encoder, enabling text‑to‑image generation within a single architecture.
Text-Centric Generation
SenseNova‑U1.5 sets new text‑centric benchmarks, topping open‑source scores across multiple datasets.
SenseNova‑U1.5 achieves the highest average score on CVTG‑2K.
Average 0.948, outperforming all other evaluated models.
**Table 11.** Quantitative evaluation results on CVTG-2K. The parameters of the generation component are denoted as # Params
**Table 12.** Quantitative evaluation results on LongText-Bench. A in # Params denotes activated parameters during inference.
**Table 13.** Quantitative evaluation results on IGenBench. The parameters of the generation component are denoted as # Params.
**Table 14.** Quantitative evaluation results on BizGenEval. Each cell reports hard / easy testset scores.
**Table 15.** Quantitative evaluation results on WISE [88]. The parameters of the generation component are denoted as # Params.
Editing and Reasoning Capabilities
We evaluate editing and reasoning performance across multiple benchmarks.
We assess general image‑editing capability on four suites and then probe reasoning‑centric editing on several specialized benchmarks.
**Table 16.** Quantitative evaluation results on ImgEdit. The parameters of the generation component are denoted as # Params.
SenseNova‑U1.5 attains the highest overall ImgEdit score.
Table 16 reports an overall value of 9.47 for SenseNova‑U1.5, outpacing the next best open‑source model.
**Table.** Quantitative evaluation results on GEdit-Bench. The parameters of the generation component are denoted.
**Table 18.** Quantitative evaluation results on the WeEdit benchmark. Note that IA, TC, and BP denote Instruction Adherence, Text Clarity, and Background Preservation, respectively. Avg. is the arithmetic mean of the three Overall scores.
SenseNova‑U1.5 achieves the top average on the WeEdit benchmark.
Table 18 lists an average of 7.46 for SenseNova‑U1.5, higher than any other open‑source entry.
**Table 19.** Quantitative evaluation results on OmniRef-Bench. We adopt both objective and MLLM-based evaluation protocols.
With positional encoding, SenseNova‑U1.5 attains the highest MLLM score on OmniRef‑Bench.
Table 19 records a MLLM score of 0.72 for SenseNova‑U1.5 (w/ PE), surpassing the open‑source mean of 0.69.
**Table 20** Quantitative evaluation results on RISEBench. The parameters of the generation component are denoted as # Params. Note that CoT denotes chain-of-thought prompting before outputting the resulting image.
Chain‑of‑thought prompting improves SenseNova‑U1.5’s overall RISEBench score.
Table 20 reports an overall of 3.63 for the CoT‑enabled variant, higher than the non‑CoT baseline.
**Table 22.** Quantitative evaluation results on VBVR-Pro-Bench. We evaluate two model settings: image generation and interleaved generation. We report both in-domain (ID) and out-of-domain (OOD) scores as percentages, together with category-wise performance.
SenseNova‑U1.5 sets new records on VBVR‑Pro‑Bench.
Table 22 lists overall scores of 67.6 (ID) and 68.9 (OOD) for SenseNova‑U1.5, the highest among both open‑source and closed‑source models.
**Table 23** Quantitative evaluation results on Uni-MMMU-GaU and RealUnify-GEU. For Uni-MMMU-GaU, we report text accuracy (T) for all tasks and sample-level accuracy for the multi-step Maze and Sliding Puzzle tasks. For RealUnify-GEU, MR, MT, AF, and CN denote mental reconstruction, mental tracking, attentional focusing, and cognitive navigation, respectively.
SenseNova‑U1.5 dominates the Uni‑MMMU‑GaU benchmark.
Table 23 shows a Uni‑MMMU‑GaU score of 86.0 for SenseNova‑U1.5, well above the next best model.
Conclusion
We wrap up the paper and list all contributors.
We present SenseNova‑U1.5 as a further step toward native multimodal systems where visual understanding and generation share a unified computational substrate.
Our extensive results show that a compact native representation can sustain strong understanding while supporting high‑fidelity generation, editing, interleaved generation, and reasoning‑intensive visual tasks.
The central opportunity of multimodal unification lies in capability interaction: reasoning improves visual creation, and generation becomes part of the model’s internal problem‑solving.
This blurs the traditional boundary between perception and creation, pointing toward foundation models that interpret, imagine, and act upon visual worlds as a unified visual intelligence.
Project Sponsor and Advisor: Dahua Lin.
Senior Project Lead: Lei Yang, Lewei Lu, Quan Wang, Ruihao Gong, Wenxiu Sun, Ziwei Liu.
Project Lead: Haiwen Diao, Jiahao Wang.
Core Contributor: Chenjing Ding, Hanming Deng, Jiangnan Chen, Ruixi Zhang, Ruohui Wang, Wenwen Tong, Xiangyu Fan, Yubo Wang, Yue Zhu, Yuwei Niu, Zhengqi Bai, Zhiqian Lin, Zhitao Yang, Zhongang Cai.
Contributor: Bo Yang, Chen Feng, Chengguang Lv, Guangjia Liu, Guanlin Wang, Hanyu Zhang, Haojia Yu, Hongcan Xiao, Hongli Wang, Huan Wu, Huaping Zhong, Jian Fang, Jianan Fan, Jiaqi Li, Jiefan Lu, Jing Zuo, Jingcheng Ni, Junxiang Xu, Linjun Dai, Mutian Xu, Peishen Yan, Penghao Wu, Ruijie Mao, Ruisi Wang, Shihao Bai, Shuang Yang, Shuya Yang, Shuyan Zheng, Silei Wu, Siying Li, Tao Chu, Tianbo Zhong, Tongxi Zhou, Weichao Luo, Weichen Fan, Wenhao Jia, Wenjie Gao, Xiangli Kong, Yan Li, Yang Yong, Zimo Wen, Zixuan Qian.
Acknowledgement: Boyu Guan, Chen Wei, Chenyang Gu, Fanzhou Wang, Haoge Deng, Houyuan Chen, Huchuan Lu, Jiaxu Li, Kaizhe Yang, Lianqiang Shi, Oscar Qian, Qingping Sun, Wanqi Yin, Wenjie Ye, Xiaotong Li, Xuehai Bai, Yijia Fan, Yue Huang, Yukun Wei, Yves Wang, and Zongpu Zhang for their support.