🅛🅣🅧 Gemma 3 Prompt Enhancer

LTX is notoriously prompt-hungry - short or vague prompts give you static, glitchy, or ignored-the-input results. This node takes your one-line idea and expands it into the long, detailed, chronological paragraph LTX-2 actually wants, using the same Gemma 3 model that already sits in your LTX-2 workflow as the text encoder. Type "a woman at a coffee shop talking on the phone," and it hands back a full cinematic description with lighting, wardrobe, an integrated soundscape, and quoted dialogue.

It's part of ComfyUI-LTXVideo, Lightricks' official pack. The enhancer is a convenience - and, honestly, a slightly controversial one (more below) - but it's built specifically around LTX-2's audio-video model, which is why its default instructions insist on describing sound and speech, not just visuals.

How it works

The node feeds your raw prompt, plus a long system prompt, into Gemma 3 as a chat completion and returns the model's rewrite. The baked-in system prompt is doing the real work: it tells Gemma to keep every element you asked for, invent concrete detail where you were vague, write in present-progressive verbs, lay events out chronologically, and - the LTX-2-specific part - describe a complete soundscape and put any speech in quotes with a described voice. It's an output node, so it shows you the enhanced text right in the graph.

The inputs that matter

• clip - the Gemma model, from the Gemma 3 Model Loader. Same encoder your workflow already loads, so no second model.
• prompt - your raw idea. Short is fine here; expanding it is the point.
• system_prompt - the instruction block. It ships with a strong default; only touch it if you want to change the house style (e.g. drop the audio emphasis).
• max_tokens - length cap on the rewrite (default 512).
• bypass_i2v - a toggle for image-to-video behavior. There's also an optional image input, so the enhancer can look at your I2V source frame and describe that, and a seed for reproducibility.

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

Installing it

ComfyUI Manager → search LTXVideo → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo then restart. It's under lightricks/LTXV. No new model to download beyond the Gemma 3 text encoder you already need for LTX-2 (the full folder in models/text_encoders/).

Where people get burned

• Trusting it blindly. The loud community advice on LTX-2 is to check what Gemma writes - drop a Show Text node on the output. People have found the auto-enhanced prompts surprisingly basic, and many get better results writing the paragraph themselves (or with an external ChatGPT/Gemini/Grok pass against Lightricks' prompting guide) and bypassing this node entirely.
• Paying the memory tax twice. Running Gemma to enhance on top of running it to encode means the 22GB model is doing double duty. On a tight card that's more offloading and more slowdown - a real reason to pre-write prompts instead.
• Unwanted embellishment. The default system prompt invents lighting, wardrobe, and sound you didn't ask for. That's a feature for a bare idea and a bug when you had a precise prompt - in that case, skip it or trim the system prompt.
• Camera moves and dialogue. The instructions are careful not to invent camera motion or dialogue unless you mention them, so if you want a specific push-in or a line of speech, say so explicitly in your raw prompt.