CheckPointLoaderSimpleFromString

This is the small one. CheckPointLoaderSimpleFromString does exactly what CheckpointLoaderSimple does, except the checkpoint name is a STRING input you can wire instead of a widget you have to click.

That's the whole node. So why does it exist? Because the moment any part of your workflow is driven by data - a TOML prompt file, an [_exports] block, a JSON blob from an API node, a note you paste at the top of the graph - the dropdown becomes the one place you still have to reach in by hand. This removes it.

How it works

It owns an instance of the core CheckpointLoaderSimple and calls load_checkpoint(ckpt_name) with whatever string arrives. No reimplementation, no caching layer of its own, so it loads precisely what the stock node loads from the same filename and it behaves the same way on a re-queue.

It's also the plain-ckpt_name twin of the pack's CheckPointLoaderFromString, which adds clip_name, vae_name and stop_at_clip_layer on top. Use this one when the checkpoint is genuinely all you need; reach for the bigger one when you're loading a separate text encoder or VAE, which is most of the 2026 models.

Inputs and outputs

One input: ckpt_name, the filename relative to models/checkpoints - with subfolders as forward slashes, e.g. zimage/model.safetensors. Three outputs: MODEL, CLIP, VAE, the same trio as the core node, so it slots into any graph that already has a checkpoint loader in it.

Install

Manager search for the pack name, or:

cd ComfyUI/custom_nodes
git clone https://github.com/morino-kumasan/comfyui-toml-prompt

and restart. There's nothing to pip install - this repo's requirements.txt is empty, and the README's install snippet is out of date (it still uses the pack's old comfyui-utils name over SSH). The pack is genuinely light: no model downloads of its own, it just orchestrates core nodes you already have.

Where people get burned

No dropdown means no autocomplete and no pre-flight check. A misspelled filename passes graph validation and blows up when you queue. If you're typing the string from memory, copy it out of the loader dropdown in a scratch workflow first.

Same rule as the core node: the file has to be where ComfyUI looks. Anything in ComfyUI/models/checkpoints (including subfolders) is fair game; a file sitting in your Downloads folder is not, no matter how you spell it.

Sharing a workflow now shares a filename, not a selection. The old dropdown at least told the recipient which of their models you meant when they opened it. A string is just a string, so a workflow built around this node arrives needing a specific file at a specific path. Fine for your own machine, worth a README if you're publishing.