LucidFlux_SM_Cond

What it is

LucidFlux_SM_Cond is the rare node in this pack you can ignore and still run the whole pipeline. It sits between LucidFlux_SM_Model and LucidFlux_SM_KSampler, and its entire job is to merge a LoRA into the loaded Flux model before sampling. There's exactly one documented use case: the FLUX.1-Turbo-Alpha distillation LoRA, which the README lists as optional and notes cuts you down to "8 steps up."

How the LoRA merge works

The mechanism is not what you might expect from ComfyUI. A normal LoRA loader produces a model patch - a wrapper that applies the delta at sampling time and can be stacked and bypassed. This node doesn't do that. It literally merges the LoRA weights into the model's state dict: for each lora_up weight it finds the matching lora_down and the original parameter, computes original + (down @ up) * scale, and copies the result back into the model. Applied, merged, done. That means there's no on/off toggle mid-graph - once this node runs, the LoRA is in the weights. Not a problem for the intended flow; just don't expect to un-apply it downstream.

The inputs:

• model: the bundle from LucidFlux_SM_Model. It reaches inside and patches the transformer.
• lora: a combo of files in models/loras/. Set to none and this node is a pure passthrough - it hands the model straight through untouched. That's your "I don't want the LoRA" switch.
• scale: 0 to 1, default 1.0. This is the merge strength. At 1.0 the turbo LoRA is fully applied; if it's over-riding the restoration quality, dial it down rather than turning it off entirely.

Output: model, which feeds the KSampler. The dual_condition_branch from the Model node rides along inside the bundle untouched.

About that LoRA: FLUX.1-Turbo-Alpha is Alibaba's step-distillation LoRA for Flux, and "8 steps" is the claim you're buying. The distillation tradeoff is the usual one - fewer steps, less refinement time, and detail quality is what gives first. The 1K example workflow runs the turbo LoRA at scale 1.0 with 20 steps, which reads a little like belt-and-suspenders; the author's own note suggests you can drop steps toward 8 when you use it. If you're on a slow card, that's the real appeal: the sampling stage is a 12B model, and cutting 20 steps to 8 is a 2.5x wall-clock win on the most expensive part of the pipeline.

Install

Install is shared with the pack: clone smthemex/ComfyUI_LucidFlux into custom_nodes (or ComfyUI Manager → search "ComfyUI_LucidFlux"), pip install -r requirements.txt, restart. The only per-node download is the LoRA itself, and it's optional - drop FLUX.1-Turbo-Alpha into models/loras/ only if you're chasing speed. And a licensing footnote that matters if you're commercial: the turbo LoRA is fine, but it rides on a FLUX.1-dev base, and Dev is non-commercial. This whole pack assumes a Dev dit, so if commercial use is on the table, this isn't the restoration stack for you.

Should you use it?

Honestly, this is the node to reach for when you want speed and the one to bypass when you don't. If your first LucidFlux render looks slightly mushy on fine detail, test the same image with the lora set to none before you blame the sampler - this is a merge, and once it's in, it's in.