VELVET VICE MiniMax H3 — Final Prompt Preview

By the time a prompt reaches the H3 sampler in the Velvet Vice workflow, it may have been through a lot. It started as your idea, got rewritten by an Ollama vision model reading your source image, possibly had duration and ending-mode direction fused in, and came out the other side as one final string. VelvetViceMiniMaxH3FinalPromptPreview is the checkpoint that shows you that exact final string on screen before the render commits - because with a 33B video model, you really don't want to discover the prompt was mangled after the sampler has been churning for ten minutes.

It's the smallest possible idea done well: an input, an output, and a UI display. The node takes a prompt (STRING, wired from the prompting pipeline rather than typed by hand), shows it in the ComfyUI preview with word and character counts, and forwards the identical string onward. The description nails it: "Displays and forwards the exact prompt delivered to the MiniMax H3 image-to-video subgraph."

What actually happens

Run it and two things land in the node's UI: the full final prompt text and a stats line like 14 words | 92 characters | exact MiniMax H3 input. The console gets the same summary. The node also returns the prompt unchanged on its prompt output, so it slots into the middle of the chain without breaking the data flow - it's a monitor you can insert anywhere without rewiring.

Two implementation details worth knowing because they explain behavior you'll notice. The node's IS_CHANGED always returns NaN, meaning it re-executes on every queue even when its input looks identical - cheap here, and it guarantees the preview always reflects the current run rather than a cached value. And it's strict about types: feed it anything that isn't a string and it raises rather than silently coercing, which is a deliberate "fail loudly at the preview, not at the sampler" choice.

Inputs and outputs

• prompt (STRING) - the assembled final prompt, force-wired input.

Output: prompt (STRING), pass-through - plus the UI preview with word/character stats.

Install and how you'd use it

Same pack, same install:

cd ComfyUI/custom_nodes
git clone https://github.com/Velvet-Vice/velvet-vice-minimax-h3

then restart.

Where this node earns its keep is debugging prompt surgery. If your vision-assisted prompt looks great in the Prompt Director panel but the render doesn't match your image, the mutation happened between director and sampler - and this node is where you'd catch it. If the audio-silencing or ending-mode framing you set in the Director isn't showing up in the final string, that's also visible here before you spend the GPU time. Some people leave it in permanently as a "last look before the expensive step" habit; that's a fine use. It's the rare preview node that can't slow you down or corrupt anything, because all it does is read text and hand it back.