SenseNova_SM_Sampler

This is the node where SenseNova-U1 actually does the work. The pack is two nodes - SenseNova_SM_Model loads the weights, this sampler runs them - and "runs them" covers four different jobs depending on what you feed it. Text-to-image. Image editing. Image question-answering. And interleaved generation, where the model writes text and draws images in the same response, like a caption-and-illustrate machine. Most samplers in ComfyUI do one thing. This one is a Swiss Army knife, which is both its appeal and the main source of confusion.

How it works

SenseNova-U1 is a native multimodal model in the NEO-Unify style: language and vision share one transformer, so understanding and generation are the same mechanism. The sampler drives it via flow-matching steps rather than classic denoising, which is why cfg defaults to a mild 1.0 instead of 7, and why there's a second knob, img_cfg, that scales how strongly the input image conditions an edit. Turn on think_mode and the model first emits a <think>...</think> reasoning block, then generates. Nothing here calls an API - it's all local, all in the weights you loaded.

The img_mode dropdown picks the job, and it's worth understanding the fallthrough in the source because it surprises people:

• t2i is implicit. No input image + any mode that isn't interleave = plain text-to-image. So img_mode set to edit with no image silently becomes T2I. Feed it an image and it's an edit.
• edit - image editing with a prompt. This is the mode the pack was built around (infographic work is the model's showcase).
• vqa - image understanding: you ask a question about the image, the text output is the answer.
• interleave - text + up to interleave_max images, in or out. This one works with or without an input image.

The inputs you'll actually touch

Most of the schema is pass-through text-decoding knobs (temperature, top_p, top_k, repetition_penalty) that only matter for the VQA/interleave text output. The ones that change your life:

• prompt - multiline, obvious.
• seed / steps - steps defaults to 8, which only works if you loaded the 8-step LoRA in the model node. Without it, expect to crank this toward 50 and it will hurt.
• target_pixels - the model trains on fixed resolution buckets, so you pick an aspect ratio, not a pixel count: 1:1 maps to 2048×2048, 16:9 to 2720×1536, and so on. No free-form resolution.
• cfg / img_cfg - guidance scale and image-guidance scale. 1.0 is a fine place to start for both; push cfg up if generations look washed out.
• prefetch_count - layer-swap offload control. Default 1 is the RAM-hungry safe mode for small cards; set 0 on 16 GB+ VRAM to disable swapping and speed things up.
• image (optional) - the IMAGE input that flips you from T2I into edit/VQA/interleave.

Outputs

• image - an IMAGE tensor, wired straight to a Save/Preview node.
• text - a STRING. In VQA mode this is the actual product. In T2I/edit mode it's the <think> block if you enabled think_mode, otherwise a placeholder like "not think mode."

One quirk worth knowing: in vqa mode the image output is literally a black placeholder tensor, because the job produced text, not pixels. Don't panic and don't save it - read the text output.

Installing it

It ships in the same pack as the loader, so you get both at once: ComfyUI Manager → search "SenseNova", or:

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

Then restart ComfyUI and grab the models (GGUF + optional 8-step LoRA) per the loader's README. Dependencies include transformers >= 4.57.1, < 4.58.0 - the upper pin matters, so don't let a newer transformers from another pack clobber it.

Common issues

The recurring ones, all grounded in the pack README and the author's own update posts:

• OOM or crash on load - check your system RAM first, not your VRAM. The model wants 36 GB+ of RAM to load via layer swapping; 8 GB VRAM is genuinely fine.
• Slow generations - you're missing the 8-step LoRA, or you're on prefetch_count swapping when you have the VRAM to disable it.
• Known 8-step LoRA artifacts - the community has flagged occasional text repetition and white-background fills with the fast LoRA. It's a tracked trade-off, not your config.
• Black "image" in VQA - not a bug, see above. The answer is in the text output.