Prompt Load - Star7

You know the feeling: you generated a MiniMax H3 clip three days ago, the node graph is gone, and all you have left is the MP4. Retyping the prompt from memory is hopeless. Prompt Load - Star7 is the recovery tool for exactly that situation - drop the file back onto the node and it digs the original prompt out of the metadata.

It ships inside the minimax-h3-chunk-star7 pack, the same one that gives you the VRAM-chunking core node for running H3 on modest cards. Don't let that scare you: this node is one of the pack's "lightweight helpers." It calls no API, needs no key, downloads no model, and doesn't touch the attention pipeline at all. It's a pure prompt-extraction utility, and it's genuinely useful even if you never use the chunk node.

How it works

The mechanism is nicer than you'd expect for a "load" node. ComfyUI embeds the workflow and prompt into output files as text metadata - PNG tEXt chunks, A1111 parameters blocks, and (for videos that were written properly) container metadata. Prompt Load reads all of it. Drop an image, video, or workflow JSON onto the node and its frontend sends the file to a backend route that:

• walks PNG text chunks, EXIF, and video metadata,
• flattens workflow JSON graphs to find every string widget,
• scores each candidate (prompt-ish node titles and text inputs score high, negative prompts get hammered down and are never auto-picked),
• and hands you the winner.

The ranking has a cute bias: longer text wins because a real H3 prompt is a wall of text while hidden helper widgets hold short labels. If it picked the wrong one, there's a "Candidates" button on the node and you can click any alternative.

The node itself is boring in the best way - one prompt input (multiline string) and one prompt output. Wire the output into your text encoder or wherever your prompt goes, then import by dragging a file onto the node. If you're mid-edit when it finishes reading, it won't clobber your text; failed or unsupported files leave your prompt untouched. One clever fallback: a video with no metadata gets checked for a same-name .png/.webp companion sitting in your input/output/temp folders.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/star7code/minimax-h3-chunk-star7.git

then restart ComfyUI. Or search "MiniMax H3 Activation Chunk - Star7" in ComfyUI Manager. There are no hard Python dependencies - but reading video metadata needs PyAV (av), which ComfyUI normally bundles. If a video drop dies with a generic parse error, suspect that first.

Where people get burned

• Re-encoded files are useless. A Reddit-hosted or screenshot version of your output has been recompressed, and that wipes the metadata. Only the original bytes carry the workflow.
• API-generated PNGs carry the flattened prompt but not the full editable workflow graph - you'll recover the prompt, not the node layout.
• For videos, the metadata has to exist in the first place. That's the other half of this story: the pack's Video and Workflow Export node writes the prompt and workflow into the MP4, so this node can read it back out. Export first, then you've got a self-documenting file.

It's a small tool, but once you start exporting H3 videos with their workflow embedded, this is the node you'll reach for every time you want to re-use a prompt from a week ago.