SenseNova_SM_Model

If someone tells you they're running SenseNova-U1 in ComfyUI, this is the node they mean - the model half of a two-node pack. SenseNova-U1 is SenseTime's attempt at a native multimodal model: one set of weights that understands images and generates them, with no separate visual encoder and no VAE bolted on. No API key, no cloud call, nothing to sign up for. You download a GGUF, point this loader at it, and the whole thing runs on your card. That's worth pausing on, because the ecosystem is mostly split between "models that draw" and "VLMs that describe." This one does both, and the loader is the door.

How it works

The node is a thin wrapper over the model's own inference code. It scans three Comfy folders - ComfyUI/models/gguf, ComfyUI/models/diffusion_models, and ComfyUI/models/loras - and gives you a dropdown for each. You pick one of the first two, and the loader builds the model with the transformers library (bf16 by default), loading a GGUF through the gguf package or a .safetensors from diffusion_models. The LoRA slot is optional and the whole point of it: the 8-step LoRA is what makes SenseNova-U1 practical on consumer GPUs (more on that below).

One thing to get straight: the model output is not a standard Comfy MODEL like you'd feed a KSampler. It's the pack's own wrapper object, and it only wires into the second node in the pack, SenseNova_SM_Sampler. Don't try to route it anywhere else.

The inputs that matter

Four dropdowns, that's the whole node:

• diffusion_models / gguf - the checkpoint. Pick exactly one. GGUF (SenseNova-U1-8B-MoT-8step-Q6_K.gguf or the smaller A3B MoE Q4_K_S) is the RAM-friendly path; the bf16 .safetensors files (including the Infographic-V3 int8/fp8 ones) are the "I have a bigger machine" path.
• lora - the SenseNova-U1-8B-MoT-LoRA-8step-V1.0.safetensors. Load it or you're looking at 50-step generations instead of 8.
• attn_backend - auto, flash, or sdpa. Leave it on auto unless you know your flash-attn build works; sdpa is the safe fallback.

Installing it

Same drill as any custom node. Via ComfyUI Manager, search "SenseNova" (or "ComfyUI_SenseNova_U1") and install. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/smthemex/ComfyUI_SenseNova_U1
pip install -r requirements.txt

Then restart ComfyUI. The requirements.txt pulls transformers >= 4.57.1, < 4.58.0 - that upper pin is real and worth knowing before you install, because a bleeding-edge transformers from another node can break this pack. If something in your env upgrades past 4.58, this loader will sulk.

Model files go in the standard spots. The pack even creates ComfyUI/models/gguf/ for you if it's missing:

ComfyUI/models/gguf/             SenseNova-U1-8B-MoT-8step-Q6_K.gguf   (or the A3B MoE Q4_K_S)
ComfyUI/models/diffusion_models/ SenseNova-U1-8B-MoT-8step-merge_bf16.safetensors  (optional)
ComfyUI/models/loras/            SenseNova-U1-8B-MoT-LoRA-8step-V1.0.safetensors

Links are in the pack README (HuggingFace smthem/SenseNova-U1-8B-MoT-Merger-gguf for GGUF, sensenova/SenseNova-U1-8B-MoT-LoRAs for the LoRA). These are multi-gigabyte downloads - a Q6_K 8B is around 6 GB, so budget the disk and the bandwidth.

Where people get burned

System RAM is the constraint nobody expects. The author's own update posts hammer this: it runs fine on 8 GB VRAM via layer swapping, but needs 36 GB+ of system RAM to load without crashing - "regardless of VRAM." If you have 16 GB or more of VRAM and a beefy card, the README recommends setting prefetch_count = 0 in the sampler to disable the layer-swap and load the whole Q6 GGUF into VRAM instead; on smaller cards, leave swapping on and let the RAM soak it. First load is also slow - state dicts get matched and layers shuffled around before the sampler ever runs. That's normal, not a hang.