🎲 Random Float in Range

Most people search for the right CFG by typing a number, queueing, typing another number, queueing. Twenty minutes, ten images, two useful data points. 🎲 Random Float in Range does that sweep for you: set 4.0 to 7.5, queue a batch, and sort the keepers afterwards. Same story for denoise on an img2img pass - 0.4 to 0.75 in one run tells you more about your image than any tutorial.

It's a decimal-roll node, and the decimals input is what makes it pleasant. Rolling 6.28491734 for a CFG value is noise you don't need.

How it works

min_value and max_value are floats - the range stretches to ±1e10, so huge bounds are fine - and if you enter them backwards the node swaps them for you. The draw is uniform across that range, inclusive at the ends in practice.

decimals rounds the result: -1 (the default) keeps full precision, 0 gives a whole number as a float, 2 gives 6.28. Rounding with 0 is a nice trick for randomising a denoise into tidy steps without reaching for the integer node.

seed defaults to -1. A value ≥ 0 makes the roll reproducible via a locally seeded RNG; -1 means "fresh each execution". Heads up on caching: the pack's random nodes only re-roll when the node actually executes, and ComfyUI skips cached nodes whose inputs didn't change. The plain Random Seed node in the pack forces a re-run with a timestamp fingerprint; this one doesn't. If your "random" value looks frozen, that's the cache - change something upstream, or set a seed and increment it.

Inputs and outputs

Two required inputs, min_value and max_value, both FLOAT; two optional ones, decimals and seed. One output, random_float, FLOAT.

The natural wiring is CFG or denoise on a KSampler: right-click the widget → Convert widget to input → connect. Both are direct float inputs, no conversion needed. Beyond that it's any float widget - a LoRA strength on a loader you've converted, an upscale factor, a mask feather radius.

Keep the range honest. Randomising CFG between 2 and 12 is not an experiment, it's a lottery; one end of that produces mud and the other produces charcoal. Between 4.5 and 6.5 is a range where every result is technically a picture and you're actually comparing something.

Install

ComfyUI Manager → search "ComfyUI-mnemic-nodes" → install → restart. Alternatively:

cd ComfyUI/custom_nodes
git clone https://github.com/MNeMoNiCuZ/ComfyUI-mnemic-nodes

Pure Python node; the pack's requirements.txt installs a heavier set (transformers, opencv-python, tiktoken, piexif) for its other nodes. No model downloads and no API keys.

Common issues

The value never changes. Cache or a fixed seed. Also check nothing downstream is being served from a cached run - swap a node's input to force recomputation.

Endless decimal places. decimals is still at -1. Set it to 1 or 2.

Wired it to a node and nothing happens. You connected to a widget that's still a widget. It must be converted to an input first, and the converted socket keeps the original type - FLOAT here, so it won't satisfy an integer input.