Conditioning (Gaussian Blur)

Conditioning Gaussian Blur takes a CONDITIONING and runs a Gaussian blur along its token dimension. That's it. One input, one slider, one output - and it does something you can't do by editing text: it makes the prompt's influence softer without removing anything.

Here's why that's useful. Your positive and negative conditioning are tensors, and each token's features tell the model what to emphasize. When you blur along the token axis with a conv1d, adjacent tokens start to agree with each other - the embedding loses its sharp per-token peaks and becomes a smoother, more averaged statement of intent. The practical effect is that the model pays less attention to any single precise detail and more to the general vibe. Lower sigma is a light touch; crank it up and the prompt stops dictating and starts suggesting.

It's a genuinely handy debugging tool for that frustrating situation where a single word is wrecking every render: instead of deleting it and losing the rest, blur the whole conditioning down and see if the direction of the problem survives. It's also a lazy-man's prompt mixing - blur two conditionings, average them, and you get a blend that no amount of comma-juggling quite produces. And if you want to do more than soften - i.e., split into frequency bands and re-weight them - that's exactly what the sibling ConditioningFrequencySplit node is for; this node's blur is the low-pass that node uses internally.

Inputs: conditioning and sigma (default 0.75, range 0–10). sigma <= 0 is a no-op that passes the conditioning through untouched. The one output is a CONDITIONING of the same shape, metadata preserved, pooled output untouched. The implementation is a per-feature-group 1D convolution, so it's effectively free even on CPU and works identically on SD, SDXL, and the newer Qwen-encoded models.

Installing

Part of the Skoogeer-Noise pack by ttulttul:

cd ComfyUI/custom_nodes
git clone https://github.com/ttulttul/Skoogeer-Noise

Restart ComfyUI or install via Manager ("Skoogeer-Noise"). Dependencies are only torch, numpy>=1.26, einops, pyyaml - nothing to download beyond the repo itself.

Gotchas

Don't expect this to act like image blur. The token dimension is short (often 77 or a few hundred tokens), and blurring it changes weighting, not legibility - the conditioning doesn't become "blurry text," it becomes "less insistent." The failure mode people hit is overshooting: a big sigma can make the prompt so mushy that the model ignores it and freewheels on the noise prior, which looks like the prompt "stopped working." Start at 0.5–1.0, and remember sigma = 0 is the escape hatch that returns your exact original conditioning.