LucidNFT_SM_Cond

Here's a subtle thing about this pack: you can run the whole chain without LucidNFT_SM_Cond and it'll work. You'd just be running LucidFlux, the base restoration model - the thing this whole project is a fine-tune of. The actual LucidNFT improvements ship as two PEFT LoRA adapters, and this small node is the only thing that merges them in. Skip it and you're not really running LucidNFT, which makes this the easiest node in the pack to quietly misuse.

How it works

LucidNFT_SM_Cond takes the pipe dict that LucidNFT_SM_Model produced and does two merges. First, it looks in models/LucidFlux/lora_dit/ and merges that adapter into the Flux transformer, and it looks in models/LucidFlux/lora_condition/ and merges that one into the dual condition branch. Those two adapters (from W2GenAI/LucidNFT on Hugging Face) are the trained LucidNFT preference-optimization itself. Second, if you've pointed it at a turbo LoRA, it merges that in too - at a strength you control.

Inputs and output

• model - the output of LucidNFT_SM_Model. This is not a plain ComfyUI MODEL; it's the pack's internal pipe object that carries the dual condition branch. If you wire in a regular checkpoint model here, nothing will merge and you'll be confused why the chain "works" but isn't LucidNFT.
• turbo_lora - combo over your loras folder. The README points at alimama-creative/FLUX.1-Turbo-Alpha (optional, "8 步起" - as few as 8 steps). Leave it at none for the full-quality path.
• scale - float, 0–1, default 1.0. This scales the turbo LoRA merge, not the LucidNFT adapters. The main adapters always merge at full strength when they're present.

The single model output feeds LucidNFT_SM_KSampler.

Setup and gotchas

Same pack install as everything here: git clone https://github.com/smthemex/ComfyUI_LucidNFT into ComfyUI/custom_nodes, pip install -r requirements.txt, restart. The lora_dit and lora_condition folders go under models/LucidFlux/ (see the model layout on the LucidNFT_SM_Model page); the turbo LoRA, if you use it, goes in models/loras/ like any other.

The main trap is expectations around speed. Merging the turbo LoRA and dropping the sampler to 8 steps is the distilled fast path - and like every distillation, it trades a bit of restoration quality for a ~2.5x speedup. The KB's own take on distilled models applies here: the full 20-step run is the quality setting, turbo is the "I'm iterating" setting. The other trap is that the node prints a cheerful "Applied turbo LoRA" even when the merge silently failed to find matching keys (it catches the exception and logs rather than dying) - so if your turbo path looks like it's doing nothing, that's not your imagination. The LucidNFT adapters are the important merge anyway.