Preview Vision Prompt

Some nodes solve a hard problem; this one solves the annoying one. Preview Vision Prompt is a pass-through display node: you feed it a STRING, it shows that text inside the node so you can actually read it on the canvas, and it hands the exact same string out the other side untouched. No saving, no file writing, no conversion. That's the whole job, and it's the node you'll grab the moment you realize you've been squinting at a truncated prompt in a widget you can't scroll.

Why you'd reach for it

The pack's prompt generators - Vision Prompt Assistant, Abliteration Vision Prompt, and especially the H3 Story Director family - all output long STRING values. When you're iterating on an H3 reference prompt, you want to read the full thing and catch a truncated sentence, a mangled <Picture 2> tag, or a scene prompt that died early, before you spend a MiniMax render on it. Preview Vision Prompt is the built-in answer: it sits in your workflow between the generator and the conditioning/sampler, displays the prompt in a resizable text box, and passes it through unchanged.

It's also an output node, which matters more than it sounds: because it terminates the chain in the UI, it's a clean inspection point while you debug a prompt pipeline without adding a Save Text-style side effect.

The one input, the one output

• prompt (input) - a STRING, usually from a generator's generated_text or scene_prompt output. It's force_input, so it can't become a plain widget value.
• prompt (output) - the identical string, passed through for whatever comes next.

That's it. One in, one out, zero loss.

Install

It ships in the same pack as everything else - ComfyUI Manager → search Vision Prompt Assistant, or:

cd ComfyUI/custom_nodes
git clone https://github.com/elgalardi/ComfyUI-VisionPromptAssistant

Restart ComfyUI and it appears under the pack's text nodes. No model files, no dependencies beyond the pack itself.

Common pattern

Generator → Preview Vision Prompt → H3 prompt input. Wire it in, check the text, and when the output looks right you can either leave the node in the chain (it's just a pass-through, it costs nothing at runtime) or bypass it. Because it never writes to disk, there's nothing to clean up afterward.

If you're running the H3 Story Director, this is also the cheap way to eyeball the scene_prompt or mode_prompt outputs before they hit the Chain Plan - worth doing at least once to see what the Director actually produced.