SenseNova-U1.5-8B-MoT-GGUF

GGUF conversion of sensenova/SenseNova-U1.5-8B-MoT for local inference.

This repo currently ships an F16 GGUF (35.1 GB). Hugging Face reports 18B params and architecture sensenova_u1.5.

File	Precision	Size	Notes
F16 GGUF	16-bit	35.1 GB	Full-precision GGUF export of the official U1.5 MoT checkpoint

• Base model: sensenova/SenseNova-U1.5-8B-MoT
• License: Apache 2.0 (same as upstream)
• Architecture: NEO-unify Mixture-of-Transformers (MoT), native unified multimodal (T2I, editing, VQA-style understanding)
• Code: OpenSenseNova/SenseNova-U1
• ComfyUI nodes: OpenSenseNova/ComfyUI-SenseNova-U1

Model description

SenseNova-U1.5-8B-MoT is SenseTime / SenseNova’s native unified multimodal model for visual creation. It is trained with NEO-unify and improves patchify layers, data mix, task formulation, prompt enhancement, and post-training versus earlier U1 releases.

Intended uses:

• Text-to-image (including native high-res / 4K-oriented generation)
• Native image editing (local edits, text edits, multi-reference, insert/replace)
• Text-heavy layouts: posters, infographics, brand assets (Chinese and English)
• Instruction-heavy generation with counts, layout, style, and spatial constraints
• Region / object control via boxes, markers, and reference images

This GGUF is a community quantization/export of the official RL checkpoint. It is not an official SenseNova upload.

Capabilities (upstream U1.5)

• Higher-quality generation: composition, color, materials, lighting, local detail
• Stronger in-image text and infographic hierarchy
• More efficient native 4K-style generation
• More reliable identity and unedited-region preservation in edits
• Better multi-constraint instruction following
• More precise visual control (boxes, markers, single/multi-image refs)

Known limitations (upstream)

• Over-emphasized detail or oversaturated color on some prompts (try lower cfg_scale)
• Errors on dense, tiny, or mixed Chinese–English text
• Imperfect counts/alignment on highly constrained layouts
• Unstable small faces, hands, limbs, and fine structures
• Drift on broad, multi-turn, or multi-reference edits

GGUF-specific notes:

• F16 is large (~35 GB). Plan disk and RAM/VRAM accordingly.
• Use a SenseNova-U1-aware loader (official Python repo or ComfyUI-SenseNova-U1). Generic llama.cpp LLM chat is not the intended path for this architecture.
• Community GGUF loaders have historically needed extra handling for some tensors; if a loader crashes, update ComfyUI-SenseNova-U1 / gguf and check node issues.

How to use

Official Python (safetensors reference; GGUF via --gguf_checkpoint when supported)

git clone https://github.com/OpenSenseNova/SenseNova-U1.git
cd SenseNova-U1
uv sync
source .venv/bin/activate
# optional: uv pip install -e ".[gguf]"

Text-to-image:

python examples/t2i/inference.py \
  --model_path sensenova/SenseNova-U1.5-8B-MoT \
  --gguf_checkpoint /path/to/SenseNova-U1.5-8B-MoT-F16.gguf \
  --prompt "A cinematic mountain lake at sunrise, realistic photography." \
  --width 2048 --height 2048 \
  --device_map auto \
  --output output.png

Image editing:

python examples/editing/inference.py \
  --model_path sensenova/SenseNova-U1.5-8B-MoT \
  --gguf_checkpoint /path/to/SenseNova-U1.5-8B-MoT-F16.gguf \
  --image input.png \
  --prompt "Change the jacket to cobalt blue. Preserve the face, pose, background, lighting, and framing." \
  --output edited.png

Typical high-quality T2I knobs used in the U1 ecosystem: cfg_scale around 4, timestep_shift around 3, tens of steps (e.g. 50). See upstream examples.

ComfyUI

1. Install ComfyUI-SenseNova-U1 (ComfyUI Manager or clone into custom_nodes).
2. Install node requirements plus gguf>=0.10.0, diffusers, accelerate, transformers.
3. Place this GGUF under ComfyUI/models/gguf/ (or map the folder in extra_model_paths.yaml under gguf).
4. Graph: SenseNova U1 Local Loader (gguf_checkpoint) → SenseNova U1 Local Text to Image (or edit node) → Save Image.
5. Restart ComfyUI after adding the file so the dropdown refreshes.

Download:

hf download NANI-Nithin/SenseNova-U1.5-8B-MoT-GGUF --local-dir ./SenseNova-U1.5-8B-MoT-GGUF

Prompting

• Simple tasks: natural language is enough.
• Complex T2I/edit: use prompt enhancement / PE recipes from the SenseNova-U1 cookbook; say explicitly what must stay unchanged.
• Fast no-GPU try: SenseNova Studio playground (upstream).

Intended use & out-of-scope

Intended: research, local prototyping, design/infographic drafts, editing experiments under Apache 2.0.

Out of scope: treating outputs as ground-truth documents; unattended production of legal/medical/financial graphics; generating harmful, deceptive, or illegal imagery. Follow the Apache 2.0 license and applicable law.

Conversion

• Source: sensenova/SenseNova-U1.5-8B-MoT (RL stage)
• Format: GGUF, architecture tag sensenova_u1.5
• Quant in this repo: F16
• Quantized / uploaded by: NANI-Nithin

If you need smaller VRAM, consider community Q8/Q4 GGUFs of U1.5; quality and loader compatibility vary.

Related models

Model	Role
sensenova/SenseNova-U1.5-8B-MoT	Official BF16 / RL weights
sensenova/SenseNova-U1.5-8B-MoT-SFT	SFT stage
sensenova/SenseNova-U1.5-8B-MoT-Preview	Preview (pre-official)

Citation

@misc{sensenova2026neounify,
  title = {NEO-unify: Building Native Multimodal Unified Models End to End},
  author = {SenseNova},
  journal = {Hugging Face blog},
  url = {https://huggingface.co/blog/sensenova/neo-unify},
  year = {2026}
}

@article{sensenova2026sensenovau1,
  title = {SenseNova-U1: Unifying Multimodal Understanding and Generation with NEO-unify Architecture},
  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},
  journal = {arXiv preprint arXiv:2605.12500},
  year = {2026}
}

Please also credit this GGUF repo if you use these files.

License

Apache License 2.0. See upstream sensenova/SenseNova-U1.5-8B-MoT. This conversion does not change the license.

Disclaimer

Not affiliated with SenseTime / SenseNova. Weights are a community GGUF export. Verify hashes, loader version, and outputs before any serious use.