LTX 2.3 Prompt Relay with a messy zombie chase scene(Prompt Relay test)

A developer has demonstrated a breakthrough in AI video consistency using LTX 2.3's Prompt Relay workflow in ComfyUI. In a deliberately challenging test—a zombie chase scene featuring full-body running, multiple pursuers, a cluttered convenience store with detailed objects, scattered chip bags, and aggressive camera movements—the workflow prevented the common failure mode where AI models lose memory halfway through generation. Typically, characters change clothes, environments morph into generic warehouses, and objects lose position. Prompt Relay kept the entire sequence clean: the woman sprints through the aisle, smashes into a shelf, and scatters chip bags while zombies maintain their pursuit.

The key innovation is a two-layer prompt structure. The Global Prompt locks the character, lighting, and environment (e.g., terrified woman, dark store, horror aesthetic). The Local Prompts dictate step-by-step action using pipe separators (e.g., "woman runs through store"). This prevents the AI from getting confused by long, messy text prompts. The method isn't perfect—long local prompts can still degrade consistency—but splitting movement into timed chunks gives precise control over both environment and action. The workflow is available on GitHub and a video tutorial explains the setup. This approach could significantly improve AI video generation for storytelling, game cinematics, and rapid prototyping.