MultiModelCheckpointIterator

MultiModelAdvancedKsampler lets you chain three checkpoints. This is the version that automates the boring part: instead of picking a second checkpoint by hand and queueing runs, you paste a whole list of checkpoints into ckpt_name2_list and each queue run uses the next one. Model 1 and model 3 stay fixed; the middle of the chain is the sweep variable.

If you've ever found yourself manually swapping a checkpoint between every queue run while testing "which anime merge works with my base?", this node removes the hands. It's the grid-search sampler - the tool for building a model comparison sheet without touching the workflow between runs.

How it works

Same chaining machinery as the Advanced sampler: optional rescaled-CFG warm-up, model 1 samples up to steps_end_first, then the selected model 2 takes over, then model 3 finishes. The difference is the iterator. ckpt_name2_list takes one checkpoint filename per line, and on each execution the node picks the next entry (cycling with % len), holding its position in node state so the next run continues down the list. Two toggles shape the sweep:

• detached_checkpoint - required for iterating at all; it's the switch that says "use the list, not a fixed model."
• random_checkpoint - when on, it picks a random entry instead of sequential. There's also reset_rng to reset the iteration index back to the start.

The chosen checkpoint is baked into the settings output (serialized as ckpt_name2), so your saver records exactly which model produced each image. Note the settings output is the only model-identity signal here - unlike the "First" variant, this node doesn't expose a used_ckpt_name string, so route the settings into MultiModelPromptSaver to keep the mapping honest.

The inputs that matter

• ckpt_name1 / ckpt_name3 - the fixed bookends of the chain (dropdowns).
• ckpt_name2_list - one checkpoint per line; the sweep list.
• detached_checkpoint / random_checkpoint / reset_rng - iteration behavior.
• load_settings - replay a full configuration from a previous settings output, exactly like the Advanced node.

Outputs: LATENT (decode as usual) and settings (the serialized config for this run).

Install and the trap

Part of comfyui-rapidfire, no extra dependencies. Install via ComfyUI Manager (search "comfyui-rapidfire") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Zeks/comfyui-rapidfire

then restart.

The trap: iteration state lives on the node instance. If you reload the workflow or restart ComfyUI, the index resets and your sweep starts over - and if the list changes length mid-sweep, % len silently skips or repeats. For reproducible sweep runs, either leave random_checkpoint on (order stops mattering) or plan to run the sweep in one sitting. Also remember this is a chaining sampler: if you only want to compare single models, a plain KSampler batch is the simpler tool, and this node only earns its complexity when model 2 is genuinely the middle of a three-model chain.