SenseNova U1.5 LoRA Loader

If you're reaching for this pack at all, odds are it's for the official 8-step LoRA - the distilled SenseNova U1.5 LoRA that takes a 50-step generation down to 8 steps at CFG 1.0. That's the difference between "nice, an infographic" and "I can iterate on this live." The LoRA Loader is the little node that makes that happen, and it works a bit differently from a standard ComfyUI LoRA loader.

A normal LoRA loader takes a checkpoint and CLIP/T5 and merges adapters into them on the graph. This one does none of that. It takes the special model output from the SenseNova U1.5 AEGIS Loader and attaches a LoRA to the SenseNova runtime itself, merging it lazily - the file isn't even touched until your first generation starts. The runtime does a shallow copy that shares the same heavyweight core (model + tokenizer) and the same ComfyUI patcher, so adding a LoRA costs almost nothing until it actually renders.

Inputs:

• model - the loader output, wired straight in.
• lora_name - populated from ComfyUI/models/loras/. For this pack you want the ComfyUI-converted file, SenseNova-U1.5-8B-MoT-LoRA-8step-ComfyUI.safetensors, which ships alongside the INT8 checkpoint in the joyfox/SenseNova-U1.5-8B-MoT-FP8 repo. The raw official file from the sensenova/SenseNova-U1.5-8B-MoT-LoRAs HF repo isn't the one this node expects.
• strength - 0–2, default 1.0. For the 8-step LoRA, 1.0 is the intended distilled configuration. The README allows 0.90–1.00 for tuning, but don't wander far; this is a distillation LoRA, not a style adapter, and dialing it down mostly just de-distills the quality.

Output is another model socket that feeds the Text-to-Image node (or Image Edit, with a caveat below). Note the pack also has auto_8step_lora_settings on the T2I node, which detects an attached LoRA and silently switches to steps=8 / cfg=1.0 - so you don't have to remember the settings, just the file.

Installing and wiring

Same pack install as the loader - ComfyUI Manager searching "SenseNova", or:

cd ComfyUI/custom_nodes
git clone https://github.com/vizart-vj/ComfyUI-SenseNova-U1.5-AEGIS

Restart, drop the converted LoRA into ComfyUI/models/loras/, and restart once more so the combo list sees it. The minimal fast workflow is Loader → LoRA Loader → Text to Image → SaveImage.

Things to know

• The 8-step LoRA is officially a Text-to-Image thing. The README calls Image Edit with it "experimental." Don't be shocked if edit results are rougher than T2I results; that's the intended release scope, not a bug in your wiring.
• The LoRA won't show in the list? You either put it in the wrong folder or started ComfyUI before placing it. loras, restart, done.
• Quality tradeoff is real and documented. The community threads on SenseNova-U1's 8-step LoRA mention occasional repeated text and the odd white background at high speed. At 8 steps you're trading a little fidelity for a 6x speedup - which for churning out layouts is usually a great deal.
• There's one subtle failure mode the pack is explicitly built to avoid: LoRA runtime cloning used to leak RAM on unload in earlier builds. The shared-core design here means unloading the patcher frees the model and tokenizer for the base runtime and every LoRA clone at once, so you shouldn't see memory creep from swapping LoRAs.