SenseNova-U1 is SenseTime's open-source unified multimodal model running natively in ComfyUI through smthemex's custom node pack. One model, five modes, no VAE, no separate text encoder — the entire pipeline collapses into a single GGUF. 8B parameters of MoT (Mixture-of-Transformers) architecture handling text-to-image, image editing, VQA, and interleaved generation from the same checkpoint.
USE 'EDIT' IMG MODE WITH REFERENCE FOR EDITING.
USE 'EDIT' IMG MODE WITH NO REFERENCE FOR TEXT TO IMAGE.
Two versions are available. The 8-step Turbo (distill) is the fast option — runs at 8 steps with cfg 1.0, lower VRAM, quick iteration. The Full Base Model is the quality option — 50 steps with cfg 4.0, more prompt adherence headroom. Do not mix these settings up. Turbo with cfg 4 will blow out, base with cfg 1 will look muddy. Match the settings to the checkpoint you downloaded.
GGUFs come in multiple quant sizes. Q6_K if you have headroom, Q8_0 for max fidelity.
Installation is straightforward. Clone the custom node into your ComfyUI custom_nodes folder:
git clone https://github.com/smthemex/ComfyUI_SenseNova_U1
Then install requirements using the portable embedded python from your ComfyUI root:
python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI_SenseNova_U1\requirements.txt
Drop the downloaded GGUF into ComfyUI/models/diffusion_models/ and restart ComfyUI. The SenseNova_SM_Model and SenseNova_SM_Sampler nodes will be available.
Workflow is three nodes: Load Image (optional) → SenseNova_SM_Model → SenseNova_SM_Sampler → Save Image. In the Model node, set diffusion_models to none, pick your GGUF, leave attn_backend on auto. In the Sampler node, set img_mode to the task you want, then dial in steps and cfg per the version you're running.
Settings cheat sheet: Turbo: cfg 1.0, steps 8 Base: cfg 4.0, steps 50 img_cfg 1.0, timestep_shift 3.0, do_sample on
VRAM tip: prefetch_count 1 for low-VRAM cards (less memory, slightly slower). Only set prefetch_count to 0 if you have 32 GB or more VRAM.
Links:
Custom Node — https://github.com/smthemex/ComfyUI_SenseNova_U1
GGUF Files (Turbo + Base) — https://huggingface.co/smthem/SenseNova-U1-8B-MoT-Merger-gguf/tree/main
Credits: Model by SenseTime (https://github.com/OpenSenseNova/SenseNova-U1), ComfyUI integration by smthemex, GGUF quantization by smthem on HuggingFace. License: Apache 2.0.
https://youtu.be/SYJhzEdN1S0?si=2kRlRp1e7R4tT5bC
Description
base and turbo 8 step model
FAQ
Comments (4)
is it supposed to take half an hour to load it the first time when i hit run?
Thanks
Notes on this image model: it performs much better with comic/cartoon-style prompts than with vintage/screen-print style prompts. Comic style asks for bold outlines and flat, saturated colors, so the output stays vibrant and clean. Vintage/screen-print style asks the model to fade colors and add texture/distress, which fights against saturation and often looks muddy or washed out. Also, writing the prompt in Chinese seems to give more accurate, detailed results than English — the model appears to have been trained heavily on Chinese-language data, so it interprets Chinese prompts more precisely, especially for style and composition details.
this is the old variant. 1.5 is out now. i just only post to civitai when i feel like it now because everyone on here is weird and rude.
my huggingface has the new ggufs for 1.5