FluxGuidanceDisable

Flux Dev doesn't have a negative prompt because the model was never trained to use one. Classifier-free guidance normally means running the model twice per step - once with your prompt, once without - and pushing the result away from the "without." Flux skips that: the correction is baked into the weights, and the "Flux guidance" slider (usually 3.5) is that baked-in correction's strength, not CFG. That's why cranking CFG on stock Flux just burns the image instead of sharpening the prompt. FluxGuidanceDisable is the node that switches the baked-in mechanism off, so a real unconditional pass - and a real negative prompt - has something to attach to.

You mostly want this alongside a de-distilled Flux checkpoint (OpenFlux, LibreFlux, or your own de-distillation experiments), or when you've built a CFG-aware pipeline around Flux and want the model to stop double-correcting. It's not something to bolt onto vanilla Flux Dev and expect magic - vanilla Dev was never trained with a real unconditional pass, so disabling guidance there just removes the thing making it work without giving you anything real to replace it. The RES4LYF README puts it plainly: set CFG to 1.0 (guidance-as-usual) on standard Flux, but on a dedistilled model, skip the Flux Guidance node entirely, use plain CLIP Text Encode, and run CFG around 3.0.

What it does and what you set

It's a model patcher: plug a model in, get a patched model out, feed that straight to your sampler. Two switches:

• disable (default true) - the on/off for the baked-in guidance path. With it on, the model behaves like a de-distilled checkpoint would for guidance purposes.
• zero_clip_L (default true) - zeroes out the CLIP-L half of Flux's conditioning. Flux runs two text encoders, T5-XXL for the heavy lifting on natural-language prompts and CLIP-L (the same small encoder from SD/SDXL) contributing a secondary, more style-and-vibe signal. Zeroing it strips that contribution out so T5 alone is driving conditioning - a common move when you want cleaner, more literal prompt-following and don't want CLIP-L's SD-era baggage muddying things once real CFG is back in play.

That's the whole node. No conditioning inputs, no numbers to tune - it's a switch you flip once per workflow, upstream of the sampler.

Installing it

RES4LYF ships as one pack, so you get every node in it including this one:

• ComfyUI Manager - search "RES4LYF," install, restart.
• Manual - from your ComfyUI folder: cd custom_nodes && git clone https://github.com/ClownsharkBatwing/RES4LYF, then cd RES4LYF && pip install -r requirements.txt (or the embedded-Python pip.exe path if you're on the portable Windows build). Restart ComfyUI afterward.

No model downloads for this node specifically - it patches whatever Flux checkpoint you've already loaded.

Where people get tripped up

Nothing looks different. If you flip disable but leave your sampler's CFG at 1.0, you won't see anything - CFG 1 means no unconditional pass runs at all regardless of what this node did upstream. Raise CFG (2.5-3.5 is the usual dedistilled-model range) for the effect to show.

Images burn at high CFG on a checkpoint that isn't actually de-distilled. Disabling the guidance path doesn't retrain the model - if the checkpoint was never taught to handle a real unconditional pass, pushing CFG up just produces the classic Flux "fried" look. This node is a prerequisite for real CFG, not a guarantee it'll behave.

Confusing this with the Flux Guidance value. "Flux guidance" (the 3.5-ish slider) and CFG are two different mechanisms that happen to share a name in casual conversation. This node turns the first one off so the second one can do something - it doesn't change what the guidance slider itself does if you're still using it elsewhere in the graph.