🎲 Load Random Checkpoint

If you have six realism checkpoints and you want to know which one actually handles your prompt, the boring answer is to swap the loader six times. The 🎲 Load Random Checkpoint node from ComfyUI-mnemic-nodes does it in one queue: you list the models, hit Queue Prompt repeatedly, and get a different checkpoint each run - same graph, no rewiring. It's a batch-variation tool, not a sampler.

How it works

The checkpoints box takes one entry per line, and each line can be four different things. A bare name like realvis is fuzzy-matched against your models/checkpoints folder. A relative path like SDXL/Realistic/ is resolved from that folder. An absolute path points straight at a file. A directory path pulls in every .ckpt and .safetensors inside it, recursively. So Realism/ on its own line is a legitimate way to say "all of these".

Matching is generous by design: it looks for a case-insensitive substring hit first, and only falls back to difflib fuzzy scoring if nothing contains your string (it gives up below a 0.3 ratio). You do not need the full filename, and you do not need the extension.

Then it builds a pool, sorts it, shuffles it once with a fixed seed, and caches it. Selection is seed // repeat_count, so the index into that pool is what changes - and that's why the author's tooltip repeats the same warning twice: control_after_generate must be set to increment or nothing ever changes.

shuffle picks the mode. False gives you the "each model gets used before any repeats" behaviour - the pool order is fixed at load, so within a session you cycle through everything. True picks at random for each index instead, which means the same checkpoint can come up twice in a row. Once the loaded checkpoint is actually in memory, consecutive runs on the same index reuse it rather than reloading the file.

Inputs and outputs that matter

Four inputs: checkpoints, seed, repeat_count, shuffle. repeat_count is the interesting one - set it to 3 and seeds 0, 1, 2 all resolve to the same model, then 3, 4, 5 move to the next. That's how you generate a few samples per checkpoint instead of one, which matters because a single image is a terrible basis for judging a model.

Outputs are the standard loader trio plus a bonus: model, clip, vae, and path - the full file path of whatever it picked. Wire path into a Save Image filename prefix (or just read it off the console) and you can tell later which model produced which image. That's the difference between a useful experiment and forty unattributed files.

Install

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

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

No API key, no model downloads. Note that the pack as a whole pulls in its full requirements.txt (transformers, opencv-python, tiktoken and friends) even if you only want this one node - those exist for its Groq and metadata nodes, not for this one.

Common issues

Nothing changes between runs. Either seed isn't incrementing, or repeat_count is bigger than the number of runs you've done. This is the number one complaint with index-based loader nodes.

"Could not resolve any valid checkpoint files." The pool came back empty. Usually a relative path that doesn't sit under models/checkpoints, or a name fuzzy-matching to nothing. The node logs its matching to the console if you enable the pack's console-logging setting, so you can see exactly what it picked.

A checkpoint you didn't want keeps winning. Fuzzy matching is substring-first, so Realistic can grab CyberRealistic and epiCRealism both. Use a directory line or a fuller path when you care.

It reloads a 6.5GB model every run. Only when the index moves. Keep repeat_count at 2–4 and let it amortise the load; that's also why the cached-pool behaviour exists.