ConditioningCutoff

If you've written a prompt like "red hair, blue dress" on an SD1.5 or SDXL model and watched the dress come out with a reddish tint, you've hit the exact problem this node exists for. The README credits the technique directly: an implementation of BlenderNeko's Cutoff, one of the earlier and still-respected answers to attribute bleeding in a single prompt, predating and complementary to full regional prompting.

What it does

CLIP conditioning isn't cleanly separated by word - attention spreads across the whole encoded prompt, which is exactly why "red" can tint something it wasn't meant to describe. Cutoff addresses this by dividing the conditioning: a target word or phrase gets its own encoding, masked so its influence is limited rather than spreading across the whole sentence, and the cutoff value controls how sharply that separation is enforced. It's a lighter-weight tool than full Attention Couple regional prompting - you're not carving up the canvas into regions, you're stopping specific words within one prompt from smearing into each other.

Inputs and outputs

Required: conds - a CONDITIONING input (your prompt already encoded). cutoff - a float, 0 to 1, default 0, controlling how strongly the cutoff is applied. Output: CONDITIONING, the same conditioning with cutoff masking applied, ready for your sampler.

Installing it

ComfyUI Manager: search "ComfyUI Prompt Control", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/asagi4/comfyui-prompt-control

No model downloads for this node - it's operating purely on conditioning tensors you've already produced.

Where people get tripped up

The one limitation the pack's own known-issues list states outright: cutoff does not work with models that use non-CLIP text encoders, like Flux. That's not a bug to chase down - the README's own words are "This might be fixable, but it's uncertain if cutoff even makes sense for those models," which is an honest acknowledgment that the whole technique is built on how CLIP conditioning works and doesn't have an obvious analogue on an LLM-encoded model. If you're running Flux, Z-Image, Anima, or anything with a general-purpose LLM text encoder, this node isn't the tool for attribute bleeding on that model - it's squarely a CLIP-era fix for a CLIP-era problem, same as most of this pack's weighting and combinator syntax.

On models where it does apply - SD1.5, SDXL, Illustrious and the rest of the CLIP-encoded lineage - the cutoff value is worth treating as a dial to tune rather than something you set once and forget. Too low and you won't see meaningfully less bleeding; push it too aggressively and you risk the target phrase's own influence weakening along with the bleed you were trying to stop. Start conservative and increase it while watching for the bleed to actually resolve.