Scheduler_CondNoise

This node perturbs your CONDITIONING with noise, and it's one of the rare nodes in this pack with an author-written description baked in - translated from the original Chinese, it lays out exactly what different strength ranges do, which is worth taking at face value rather than guessing:

• 0.1–5.0: light style/detail perturbation
• 5.0–20.0: medium style/composition adjustment
• 20.0–50.0: heavy style/abstraction transformation

That's a genuinely useful map, because "add noise to conditioning" sounds like one knob but behaves like three different tools depending on where you sit on the range - subtle texture variation at the low end, real compositional drift in the middle, and something closer to abstract reinterpretation at the top. If you're chasing prompt-adjacent variety on repeated generations from the same text - the same idea behind seed variance, but applied to the conditioning tensor itself rather than the initial noise - this is a direct way to get it, tunable by exactly how much you want to preserve versus deviate.

The inputs and outputs that matter

• conditioning (required, CONDITIONING) - whatever positive or negative conditioning you want to perturb. Wire in the output of your text encoder (or another conditioning node) here.
• strength (required, default 20, range 0–50) - the noise magnitude, per the ranges above. Default sits in the "medium" band, which is a reasonable starting point if you want the effect to actually be visible without abandoning the prompt entirely.
• start_step_percent / end_step_percent (required, default 0 and 100, both 0–100) - the window of the denoising process the noise applies over. At the defaults it's active the whole run; narrow the window to concentrate the effect where it matters most for what you're after - noise applied early biases composition, noise applied late tends to read more as texture/detail variance since composition has already mostly settled by then.
• Output: CONDITIONING - feed it straight into your sampler in place of the original conditioning.

How to install it

Search ComfyUI-Apt_Preset in ComfyUI Manager, or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/cardenluo/ComfyUI-Apt_Preset.git

install.bat handles Windows dependencies; on Linux or Mac, open it, read the pip commands, and run them yourself, or let ComfyUI Manager resolve anything missing on first load. No model files involved - this operates purely on the conditioning tensor.

Common issues & troubleshooting

Prompt adherence collapses entirely. You're almost certainly in the 20–50 "heavy/abstraction" range and the model isn't holding the original description together anymore. That's not a malfunction - it's what the author's own description says that range does. Pull strength down toward single digits if you wanted variety, not reinterpretation.

No visible effect at the default. Check start_step_percent/end_step_percent haven't been narrowed to a tiny window elsewhere in a saved workflow you're building on - a 0–100 range is your safest bet while you're first calibrating the effect, then narrow it once you know what strength you want.

Trying to use this for the same kind of variation seed-cycling gives you, but more controllable. That's exactly the use case this node targets - low strength (1–5) gives you prompt-adjacent variation without the compositional lottery of a fresh seed, since the base conditioning direction is preserved and only lightly perturbed.

Applying this to both positive and negative conditioning. Nothing stops you from running two separate Scheduler_CondNoise nodes, one per conditioning branch - just be aware perturbing both at once compounds the drift, so start with one side at a time to see which is actually responsible for a given change in output.