Prompting Guide with LTX-2.3

Lightricks has released a new prompting guide for its LTX-2.3 AI video model, revealing a fundamental shift in how users should interact with the system. The core update is a larger, more capable text connector that interprets complex prompts with greater accuracy, especially those containing multiple subjects, spatial relationships, and detailed actions. This means the previous strategy of simplifying prompts for consistency is now obsolete; LTX-2.3 rewards specificity. For example, instead of prompting for 'a woman in a café,' users are encouraged to write detailed scenes like 'A woman in her 30s sits by the window of a small Parisian café. Rain runs down the glass behind her.' This level of detail reduces creative drift and yields more precise outputs.

Beyond text understanding, the model features a rebuilt latent space and updated VAE for sharper fine detail across resolutions, allowing users to effectively describe textures like fabric types and hair strands. A major focus is on motion generation: LTX-2.3 reduces freezing and produces more natural movement, but requires clear, verb-driven prompts (e.g., 'The camera slowly pushes forward as the subject turns their head') instead of vague descriptions. The update also includes native vertical video support up to 1080x1920, trained on portrait-format data, and an improved vocoder for more reliable audio alignment. The guide essentially trains users to 'block the scene like a director,' using explicit spatial and action commands to fully leverage the model's enhanced capabilities.