SenseNova-U1.5-8B-MoT is our latest native unified multimodal checkpoint for more accurate, consistent, reliable, and aesthetically compelling visual creation. Built on NEO-unify, it strengthens the patchify layers, data quality and distribution, task formulation, prompt enhancement, and post-training pipeline.
The official release focuses on six user-visible improvements:
Higher-quality image generation: improved composition and color harmony, with more realistic material rendering, natural lighting, stronger visual fidelity, and finer local details.
Better text rendering and infographic generation: more legible Chinese and English text, with clearer information hierarchy in posters, infographics, brand assets, and other text-dense designs.
More efficient native 4K generation: more coherent global structure, color harmony, and stable high-resolution output with improved generation efficiency.
More reliable native image editing: stronger preservation of subject identity and unedited content across local, text, multi-reference, insertion, and replacement edits.
Stronger complex-instruction following: more consistent execution of object counts, spatial relationships, layouts, styles, and multiple constraints within a single request.
More precise visual control: more accurate region- and object-level control through bounding boxes, visual markers, and single- or multi-image references.
The upstream environment uses Python 3.11, PyTorch 2.8, and CUDA 12.8. For other CUDA versions and optional FlashAttention configurations, see the installation guide.
Text-to-Image
bash
1python examples/t2i/inference.py \2 --model_path sensenova/SenseNova-U1.5-8B-MoT \3 --prompt "A cinematic mountain lake at sunrise, realistic photography."\4 --width 2048 --height 2048\5 --device_map auto \6 --output output.png
Image Editing
bash
1python examples/editing/inference.py \2 --model_path sensenova/SenseNova-U1.5-8B-MoT \3 --image input.png \4 --prompt "Change the jacket to cobalt blue. Preserve the face, pose, background, lighting, and framing."\5 --output edited.png
See the inference examples for more options, supported resolutions, and batch processing.
Best Practices
Direct natural-language prompts work well for clear tasks with few constraints. For complex generation or editing, use prompt enhancement when additional planning is needed and explicitly specify what should remain unchanged.
See the SenseNova-U1.5 Cookbook for setup instructions and optional Image PE, Caption-to-Prompt, and Editing PE recipes.
🌐 Use with SenseNova-Studio
The fastest way to experience SenseNova-U1.5 is through SenseNova-Studio — a 🆓 free online playground where you can try the model directly in your browser, no installation or GPU required.
Ongoing Improvements
The official release improves upon the Preview, though challenges remain in:
Over-emphasized details or colors: some prompts may produce excessive high-frequency detail or oversaturated colors, which can often be mitigated by lowering cfg_scale.
Dense text errors: dense, lengthy, small, or mixed Chinese-English text may contain errors.
Constrained layouts: exact counts, alignment, or hierarchy may be imperfect in highly constrained layouts.
Unstable human details: small faces, hands, limbs, and fine-grained object structures may remain unstable.
Complex editing drift: broad, multi-turn, or multi-reference edits may drift, especially when many regions must be preserved simultaneously.
If this project is helpful for your research, please consider starring the repository and citing:
bibtex
1@misc{sensenova2026neounify,
2 title = {NEO-unify: Building Native Multimodal Unified Models End to End},
3 author = {SenseNova},
4 journal = {Hugging Face blog},
5 url = {https://huggingface.co/blog/sensenova/neo-unify},
6 year = {2026}
7}
89@article{sensenova2026sensenovau1,
10 title = {SenseNova-U1: Unifying Multimodal Understanding and Generation with NEO-unify Architecture},
11 author = {Diao, Haiwen and Wu, Penghao and Deng, Hanming and Wang, Jiahao and Bai, Shihao and Wu, Silei and Fan, Weichen and Ye, Wenjie and Tong, Wenwen and Fan, Xiangyu and others},
12 journal = {arXiv preprint arXiv:2605.12500},
13 year = {2026}
14}
