Descriptions:
Veteran AI introduces Prompt Relay, a plug-and-play framework that solves a persistent limitation in AI video generation: most models fail to correctly execute sequences of multiple distinct actions within a single long video, a problem the creators call “semantic entanglement.” Rather than requiring a new model, Prompt Relay works alongside existing generators — including Wan 2.2 and LTX 2.3 — by precisely mapping text instructions to corresponding time segments in the output video.
The technology uses three core mechanisms: cross-attention temporal routing (preventing text instructions from leaking into unintended segments), boundary attention decay (a soft Gaussian penalty that creates smooth visual transitions between scene segments), and a global-plus-local prompt structure (global prompts define consistent background and character appearance; local prompts handle per-segment actions at specific timeframes). The video demonstrates the full implementation using a Kijai-developed ComfyUI extension, deployed via the RunningHub online platform.
A working example generates a three-action sequence — a woman releasing a bitten hoodie collar, performing a 360-degree spin, and sprinting forward — with high fidelity to each described action in order. The creator then stress-tests Prompt Relay across different camera angles and output resolutions, concluding it generalizes robustly across configurations. This is a practical, workflow-level tutorial for ComfyUI users who want to produce coherent long-form AI video with multiple sequential events, using LTX 2.3 and the Kijai extension as the specific toolchain.
📺 Source: Veteran AI · Published April 22, 2026
🏷️ Format: Tutorial Demo
