VELVET VICE LTX — Final Prompt Preview

The scariest moment in an LLM-assisted workflow is the one where you trust the LLM. Your rough idea went through a local model, came back as a long structured prompt, and you have no idea what it actually said until you've spent minutes rendering the clip. VelvetViceLTXFinalPromptPreview is the pack's answer: a pass-through node that displays the fully validated final prompt - after the Ollama release step, right before it reaches the LTX director - so you can read exactly what's about to be encoded.

What it does

One input, one output: prompt in, prompt out. The text passes through byte-for-byte - the description is explicit that this is the exact same text delivered to the LTX director, not a copy or a summary. What it adds is visibility: it shows the prompt in the UI along with a stats line - word count, character count, and the note that this is the exact LTX director input. It also prints those numbers to the console.

IS_CHANGED returns NaN, so it always re-executes and the displayed prompt is always fresh - no stale cached text from a previous run.

Why you'd want it

LTX is famously prompt-sensitive: it punishes short prompts and rewards long, specific prose, which is precisely why the Velvet Vice workflow routes your idea through a local LLM (via Ollama) before sampling. But an LLM rewrite has two failure modes the KB's LLM-in-ComfyUI essay calls out: dirty output (role delimiters and preamble leaking into the prompt) and subject drift (the enhancer inventing content you didn't ask for). This node is the checkpoint where you catch both - you see the final text before committing a render to it. If the model decided the scene needed a detail you never mentioned, you notice here and stop the queue.

Where it sits

In the Velvet Vice pipeline it comes after the Ollama release barrier (so the LLM's memory is freed before the LTX model loads - part of the pack's prompt-first strategy) and directly before the LTX director / sampler. Its output is also a plain STRING, so you can wire it into a text display node or anywhere else a prompt string is expected.

Inputs and outputs

• prompt - STRING, forced input. Feed it from the Ollama Release Barrier or the Prompt Director's final_prompt_preview output.
• Output: prompt, the identical STRING.

Installing it

Ships in VELVET VICE - LTX (velvet-vice-ltx) - ComfyUI Manager, or:

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

Restart and hard-refresh with Ctrl+F5. No pip deps.

Troubleshooting

• The prompt shown differs from what you typed - that's the point. You're seeing the post-LLM, validated prompt. If it's wrong, fix it upstream (mode, short_idea, or the LLM model), not here.
• Non-string input error - it type-checks and refuses to pass through a non-string. Reconnect the wire.
• "word 0" - an empty prompt got through; the pipeline upstream should have caught it, but if you see this, the manual_prompt was empty and a manual mode was selected.

One honest caveat: on a long clip a big prompt is a wall of text, and reading it every render gets old. But the first time it catches the LLM silently writing something you'd never approve, it pays for itself. It's a five-dollar safety mirror on a pipeline that's otherwise a black box.