SenseNova U1.5 8-Step LoRA (Final only)

SenseNova U1.5 at 50 steps is a 5–8 minute affair on a 24 GB card. SenseNovaU15EightStepLoRA is the pack's speed hack: it applies the official 8-step LoRA so you can generate in 8 steps instead of 50 - the author measured a 2048×2048 text-heavy image at about 86 seconds versus roughly 300–500 for the 50-step path. That's the difference between "I'll iterate" and "I'll wait for it."

If that pattern sounds familiar, it should: this is a step-distilled LoRA in the Lightning/Turbo lineage the KB covers - train a student to collapse the trajectory into a few big jumps. Same deal, same tradeoffs. Fewer steps means less refinement time, which is why the rules change: on this LoRA you run CFG 1 (not 4), 8 steps, and you accept the small quality haircut in exchange for speed.

What it is and isn't

It applies the official SenseNova-U1.5-8B-MoT-LoRA-8step (rank 128, alpha 8) to a Final base model. Not SFT, not Preview - the node checks the base model and gives you a clear error on the wrong combination, because the official LoRA only works with Final. It's also text-to-image only. The README is explicit: keep using the 50-step workflow for image editing. Speed adapters are bad editors.

The file goes in ComfyUI/models/loras/ and shows up in the lora_name dropdown. Make sure you grab the -ComfyUI converted file (~815 MB), not the raw official LoRA - the conversion only adds the diffusion_model. key prefix ComfyUI needs; the tensor data is byte-for-byte identical to the official release, and the node verifies that (exact file size + SHA256, 882 tensors, 294 modules, BF16) before it applies anything.

Inputs

• model - must come from SenseNovaU15Loader with a Final checkpoint. The node refuses SFT and refuses to stack the LoRA twice.
• lora_name - pick the -ComfyUI file from models/loras/.
• strength_model - default 1, which is the official value. Leave it there.

Output is a single MODEL, and then the chain continues normally: Loader (Final) → 8-Step LoRA → SenseNova Sampling Options → KSampler.

The settings that make it work

The official recipe, straight from the README:

LoRA strength: 1
steps: 8
CFG: 1
cfg_norm: none
shift: 3
sampler: euler
scheduler: normal
denoise: 1

Note the CFG: 1 - on a step-distilled adapter, guidance is effectively baked in, and running text CFG like you would on the base (4, 5, 7) overcooks it and burns the speed advantage. This is the exact "distilled models want different CFG" trap the KB warns about; it's not optional here.

Why the guardrails exist: the author has seen people wire the raw LoRA, the wrong base, or double-apply it, and get garbage. This node exists to make the failure loud instead of silent - you get a ValueError naming the exact mismatch, and the fix is usually "download the -ComfyUI file" or "use Final."

Install

Same pack, same story: search "SenseNova U1.5 (T8)" in ComfyUI Manager or git clone into custom_nodes/, restart. No extra Python dependencies. The LoRA file itself is a manual download from the README's Hugging Face link (t8star/SenseNova-U1.5-Comfy) - Manager won't fetch it for you.