🎲 Random Int in Range

You want a batch where the step count wanders between 18 and 34 instead of being nailed to 24. Or a batching setup where each run picks a different batch_size. 🎲 Random Int in Range is the boring workhorse for that: give it a min and a max, it gives you an integer inside those bounds, inclusive, every execution.

Paired with the rest of the pack, it's how you get settings variance without hand-editing widgets between queues. Randomising settings is a genuinely underrated trick - the wildcard folks randomise prompts and then run every one at CFG 7 and 20 steps, which means every twelve-of-twenty images has the same flat, overcooked look because the settings never varied.

How it works

min_value and max_value are inclusive integers and the range is huge - the inputs go to ±2^64, so you can roll a seed-sized value if you want. If you accidentally put the bigger number first, don't worry: the node swaps them for you rather than erroring or returning an empty range.

seed is optional and defaults to -1. Give it a real value and it rolls from a seeded generator, so the result is reproducible; leave it at -1 and it draws from Python's unseeded RNG. As with the rest of the pack's random nodes, -1 means "different each time it executes", and ComfyUI's cache decides whether it executes - if nothing upstream changed, the value can be reused from cache. The pack's plain Random Seed node forces a re-run with a time.time() fingerprint; this one doesn't. If you need guaranteed variance per queue, drive the input or set a seed and increment it.

Inputs and outputs

Two required inputs, min_value (default 0) and max_value (default 100), one optional seed. A single output, random_int, typed INT.

Where it actually earns its place: convert a widget to an input and feed it. Steps on the KSampler is the obvious one, but any integer widget works - batch_size on an Empty Latent, a start/end step pair for a partial denoise, a blur radius, a tile size in a tiling upscale.

Be sensible about the range, though. Rolling steps between 5 and 60 in one batch gives you six unusable images and four good ones, and you can't tell which is which without reading the metadata. Between 20 and 30 is a test you can actually learn something from.

Install

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

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

The node itself is pure Python; the pack's other nodes are what pull in transformers, opencv-python and friends via requirements.txt. No models, no API keys, nothing to download.

Common issues

Same number every run. Cache, or a fixed seed with unchanged inputs. Nudge something upstream or increment the seed.

The receiving node won't accept the wire. It needs an INT input, not a FLOAT one. Widgets converted to inputs keep their type - a CFG widget is a float, so feed that from the float node instead.

Values that make no sense downstream. Integer division is doing it. height // 8 on an odd number is silent rounding, so if you're rolling resolution, stick to multiples of 8 or 64 and save yourself a debugging session.