🅛🅣🅧 LTXV Prompt Enhancer

LTXV Prompt Enhancer is the pack's general-purpose "make my prompt longer and better" node. You feed it a short description, it runs that through a language model, and out comes a fleshed-out, LTX-friendly prompt. If you've noticed that LTX turns your two-line idea into a static or incoherent clip, this is one answer - LTX genuinely rewards long, specific prompts, and this automates the padding.

It's part of ComfyUI-LTXVideo, Lightricks' official pack. Note it's the sibling of the Gemma 3 Prompt Enhancer: this one takes its own enhancer model object rather than reusing your Gemma text encoder, so it's the enhancer you'll see across the broader LTX line, and it can also read a reference image to tailor the description.

How it works

The node hands your text (and optionally an image) to a prompt-enhancer LLM and returns its rewrite. The reference-image path is the useful twist: give it your image-to-video source frame and it describes that scene rather than guessing, which is the well-worn "have a vision model caption your input, then generate from the caption" workflow the LTX community adopted early. A token cap keeps the output from ballooning.

The inputs that matter

• prompt - your raw idea. Terse is fine; expanding it is the whole job.
• prompt_enhancer - the enhancer model object, supplied by its loader. This is what distinguishes this node from the Gemma variant - it wants a dedicated LTXV_PROMPT_ENHANCER, not your text encoder.
• max_resulting_tokens - length ceiling on the rewrite (default 256, up to 512).
• image_prompt (optional) - a reference IMAGE. Connect your I2V source here and the enhancer describes the actual frame instead of inventing a scene.

The single output is str (a STRING) - route it into your text-encode/conditioning node in place of the raw prompt.

Installing it

ComfyUI Manager → Install Custom Nodes → search LTXVideo → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo and restart. It lands under lightricks/LTXV. The enhancer needs its model weights available for the prompt_enhancer loader to pick up - like the rest of the pack, expect a sizable first-use download and keep disk headroom free.

Where people get burned

• Empty prompt_enhancer picker. This node won't run without an enhancer model wired in, and if the loader's dropdown is blank the weights aren't where ComfyUI expects them. That missing-model step trips up more setups than the node logic itself.
• Enhancement isn't a quality guarantee. The community's repeated advice on LTX is to read the enhanced text before trusting it - an auto-expanded prompt can be generic. If output disappoints, inspect the string with a Show Text node, or write the prompt yourself; hand-written long prompts often beat the auto ones.
• Gemma vs this one. On LTX-2 specifically, the Gemma 3 Prompt Enhancer reuses the encoder you already loaded and is audio-aware. If you're on an LTX-2/2.3 audio-video graph, that's usually the better fit; this general enhancer shines when you want an image-conditioned caption or you're not on the Gemma path.
• Still garbage-in. No enhancer rescues a genuinely empty idea. It amplifies detail; it doesn't supply intent - give it something concrete to work with.