LTX 2.3 is a finicky prompter. It's also the most controllable open-source video model out there once you understand how it weighs language. This guide covers what actually works — prompt structure, negative prompts that fix common artifacts, dialogue tags for lipsync, and the specific phrases that produce sharper, more on-target output.

The fastest way to test these patterns: our VirtuaVixen Studio runs every LTX 2.3 workflow free in your browser — paste a prompt, hit generate, see what the model does. For local users, the ComfyUI Workflow Pack ships with our production prompts as defaults you can mod. Discord for prompt feedback.

The Five Prompt Sections

LTX 2.3 was trained with structured tags. The Cinema-grade workflows we ship use this five-section template. You don't need all five every time, but the structure helps the model.

• Trigger words — LoRA activation keywords (e.g. k3nk4llin0n3) at the very top.
• [VISUAL]: — what's happening on screen. Camera, lighting, environment, action.
• [SPEECH]: — exact words the character says. Lipsync follows this.
• [SOUNDS]: — ambient and foley cues. Moaning, breathing, environmental sound.
• Style modifiers — “natural light”, “shallow depth of field”, “english native”, etc.