Random Checkpoint Loader (LoraManager)

Every ComfyUI user hits the same wall eventually: you've got forty checkpoints and one really good prompt, and you're manually swapping ckpt_name in the loader, running, comparing, swapping again. Tedious doesn't cover it. Random Checkpoint Loader (LoraManager) is the lazy fix - a drop-in replacement for the standard Load Checkpoint that can pick a random checkpoint from your whole pool on every run, optionally filtered by base model.

It's a genuinely weird node to reach for until you understand the two things it's actually good at. First, discovering: when you're not sure which of your models handles a prompt best, rolling through them at random surfaces candidates you'd never have bothered testing. Second, honesty-checking a prompt: if a workflow only works on one specific checkpoint, random loading finds out fast. And when it picks something you love, the model_name output tells you exactly what got loaded.

How it works

The node pulls its checkpoint list from the LoRA Manager scanner cache - the same index the manager UI browses - which means it sees your standard ComfyUI folders and any extra model folders you've registered in the manager. That extra-folder reach is a real upgrade over the stock loader.

The interesting mechanism is IS_CHANGED. ComfyUI caches node results and skips re-running when the widget values haven't changed - which would break "random on every run" instantly, since the widgets don't change. So when select_at_random is on, the node returns a non-cacheable value and forces a fresh random pick every queue run. Flip it off and it behaves like a normal loader.

Three inputs, and only two really matter:

• select_at_random - the toggle. On = ignore ckpt_name and pick randomly each run; off = load whatever's in ckpt_name.
• base_model - restricts the random pool to one base model, e.g. "SDXL" or "SD 1.5". This is the filter that keeps a Pony checkpoint from landing on an Illustrious-only workflow. Any uses everything.
• ckpt_name - the explicit pick, used when randomization is off.

Outputs are MODEL, CLIP, VAE (wire them exactly like a stock Load Checkpoint) plus model_name - a STRING with the loaded checkpoint's name. That string is the secret sauce: feed it into prompt text, the pack's metadata, or a recipe so you always know which model produced a given image.

Install

This is one node inside willmiao/ComfyUI-Lora-Manager. Via ComfyUI Manager: Manager → Custom Node Manager → search lora-manager → Install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/willmiao/ComfyUI-Lora-Manager
cd ComfyUI-Lora-Manager
pip install -r requirements.txt
# restart ComfyUI

The pack's Python deps are light; nothing heavy downloads at install. One thing to know: the checkpoint list comes from the manager's scanner, so your first launch needs to finish indexing before the random pool is meaningful.

Where people get burned

• Random model + random seed = unreproducible. If you're A/B testing, lock the seed (or use a Seed node) so a random checkpoint swap is the only variable. Otherwise you'll spend an evening not knowing whether the checkpoint or the seed changed the image.
• "value not in list" / "not found in cache." The node only offers checkpoints that still exist on disk and are indexed - a freshly downloaded model may not be in the scanner cache yet, so the random pool can lag reality. Refresh or restart ComfyUI after adding models.
• Empty pool for a base model. Pick a base_model nobody in your collection matches and you get a clean FileNotFoundError telling you to change the filter or disable randomization. That's by design - it'd rather error than silently load wrong.
• Randomizing in a shared recipe. A workflow you share with select_at_random on will behave differently for whoever opens it. Great for you, confusing for a collaborator. Use it deliberately.