LTX2 Prompt Generator 🐑

LTX2PromptGenerator takes a short input and has an LLM expand it into a detailed prompt aimed at LTX-2, Lightricks' video model. Video models are unusually punishing about prompt length, and this node exists to stop you from feeding LTX-2 the terse one-liner it hates.

It's one of the model-specific prompt nodes in ComfyUI-MieNodes, MieMieeeee's utility pack. As with its siblings, it doesn't generate video and it doesn't contain an LLM - it calls out to whichever LLM service you connect, with a system prompt shaped for LTX-2.

Why LTX-2 needs the help

LTX is the speed tier of local video: where Wan takes minutes, LTX takes seconds, and LTX-2 (the January 2026 release) bet the family on synchronized audio-video and found its niche in lipsync and music-video work (KB panel). But that same panel names the catch outright - LTX punishes short prompts. A bare "a cat walking" gives you mush; the model wants the scene, the motion, the camera, the lighting all spelled out. Describing motion over time is genuinely harder to write than a still-image prompt, and most people don't bother, which is exactly why their LTX clips look worse than the demos. This node offloads that writing to an LLM: hand it the gist, get back a fleshed-out, motion-aware prompt. The prompt-engineering principle - say more, usefully - is doubly true for video.

How it works

Simple node, three inputs:

• llm_service_connector (required) - from a Set*LLMServiceConnector node. This is the LLM that does the expansion.
• input_text (required, multiline) - your rough idea.
• seed (required) - reproducibility and variation. Same seed and input reproduce the rewrite; change it to get a different take.

The single output is ltx2_prompt (STRING). Feed it into your LTX-2 text-encode, or read it first with a ShowAnything node - worth doing on video, where a render is expensive and you want to sanity-check the prompt before committing.

Installing it

Install the pack. ComfyUI Manager → search ComfyUI-MieNodes (ComfyUI_MieNodes) → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/MieMieeeee/ComfyUI-MieNodes

then restart. No model download for the pack. Nodes appear under the sheep-emoji 🐑 MieNodes menu.

The API key step

Like every LLM node in the pack, this needs an LLM behind it, which means an API key on the connector. Type it into the connector's api_token, or drop it into mie_llm_keys.json (copy the .example first) in the plugin folder. ZhiPu and SiliconFlow have free-tier models if you just want to try it - check their sites for the current free model id.

Common issues

Errors on run usually mean the connector, not this node. Missing or bad API key, or a config_key pointing at an empty JSON entry. Isolate it with CheckLLMServiceConnectivity.

A good prompt doesn't fix the model's ceiling. The KB is honest that LTX still loses to Wan on faces, bodies, and complex motion, holding its own mainly on environments and camera moves. A better prompt gets you the best LTX can do - it won't turn LTX into Wan. If faces are your priority, that's a model choice, not a prompt one.

It's tuned for LTX-2. For images, reach for ZImagePromptGenerator or Krea2PromptGenerator; for a model-agnostic rewrite, the general PromptGenerator.