CheckpointLoaderRandom

CheckpointLoaderRandom does exactly what the name says: picks a random checkpoint out of your models/checkpoints folder and loads it. Same three outputs as the stock loader - MODEL, CLIP, VAE - which means it drops straight into any normal sampling graph as a swap-in. The point is that one node replaces "I keep manually clicking the dropdown to the next checkpoint" when you're testing which of your forty models does best on a prompt.

There's a wrinkle worth understanding before you wire it up, because it's not fully random every call. The node picks a checkpoint at random, then reuses that same checkpoint for the next N loads before picking again - that N is the max input. In practice that means: set max low (say 1) and every generated image uses a freshly-rolled checkpoint. Set it high and one checkpoint gets loaded repeatedly while you iterate on everything else, so a bad draw doesn't churn your whole batch. It's a batching device, not a dice roller.

Inputs and outputs

• seed (INT) - declared, but worth knowing: the source accepts it and never actually seeds the random choice, so don't rely on it for reproducibility. Just bump it to nudge the pick.
• max (INT, default 10) - how many consecutive loads reuse the same checkpoint before it re-rolls. The one setting that actually shapes behavior.

Outputs: MODEL, CLIP, VAE, identical in kind to the standard CheckpointLoaderSimple outputs. VAE comes from load_checkpoint_guess_config with both output_vae and output_clip enabled, so a fully self-contained graph can be built off this node alone.

Install and the honest verdict

Install is the usual clone-and-restart; ComfyUI Manager finds it by searching "Lilly":

cd ComfyUI/custom_nodes
git clone https://github.com/lilly1987/ComfyUI_node_Lilly

Restart, done. No models or extra dependencies beyond the pack's auto-installed rich/chardet.

The honest verdict: this node is a relic of early-2023 ComfyUI, when people were experimenting with "same prompt, many checkpoints" loops before XY-plot nodes and batch schedulers got good. The pack is unmaintained and the author now points to Impact Pack. If you want this exact behavior in a maintained form, the more common pattern today is a batch-iteration approach that cycles through a list of models - which people built on this pack's other loader, CheckpointLoaderSimpleText, back when this was the only game in town. It still runs, but for new work the roulette is more novelty than necessity.