🅛🅣🅧 LTXV Prompt Enhancer Loader

LTX has a well-earned reputation for punishing lazy prompts. Type "a cat on a couch" and you'll get static, chaotic, or just wrong. The model wants long, specific, detailed prose - the community's blunt summary is that "the secret sauce for LTX seems to be long prompts." This node loads the machinery that writes those prompts for you, locally, so you don't have to hand-craft a paragraph every time or paste your image into ChatGPT.

How it works

It's a loader, so it does one job: download and initialize two small models from Hugging Face and hand them off as a single LTXV_PROMPT_ENHANCER object. One is a language model that expands your short prompt into detailed video-ready text; the other is an image captioner, so if you're doing image-to-video it can look at your input frame and describe it before enhancing. That paired setup - caption the image, then write a matching prompt - is exactly the workflow people were rigging up by hand with MiniCPM or Qwen-VL back in the 0.9.x days. This just bundles it into one node.

The output feeds a separate enhancer-apply node; this loader only stands the models up.

The inputs that matter

There are only two, and the defaults are sensible:

• llm_name (default unsloth/Llama-3.2-3B-Instruct) - the small LLM that does the prompt writing. 3B is deliberately lightweight so it doesn't blow your VRAM budget on top of the video model.
• image_captioner_name (default MiaoshouAI/Florence-2-large-PromptGen-v2.0) - the captioner for I2V. Florence-2 PromptGen is a solid, small captioning model tuned for exactly this.

Leave both on their defaults unless you have a specific reason to swap them.

Installing it

• ComfyUI Manager - search LTXVideo, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo, then restart.

Heads up on downloads: the first run pulls Llama-3.2-3B and Florence-2 from Hugging Face, so expect a wait and a few gigs of disk the first time. They're small next to the video model itself, but they're not free.

Common issues

First run hangs "forever." It's downloading the two models. Watch the console - it's not frozen, it's fetching Llama and Florence.

The enhanced prompt is generic or bland. The built-in enhancers on LTX have a reputation for producing basic text - a good sanity check is to route the output through a Show Text node and actually read it. If it's weak, this is a floor, not a ceiling: write your own detailed prompt, or use a bigger external LLM. The enhancer is a convenience, not a mandate.

It ate your VRAM. A 3B LLM plus a captioner isn't nothing when the video model already wants most of your card. If you're tight, generate the prompt in one pass, save the text, and bypass the loader for the actual video run - or offload prompt-writing to an external tool entirely.