Descriptions:
Veteran AI tests a Reddit-sourced “beat-based” prompt formula for generating high-intensity dynamic motion in AI video, applying it across both Wan 2.2 and LTX Video using identical prompts and reference images. The formula structures a 5-second video as timestamped action beats—for example, “0–1.5s: man points at viewer; 1.5–2s: stands up; 3–4s: runs toward camera; 4–5s: dives to ground”—followed by cinematography notes and quality descriptors like “4K, cinematic lighting, clear texture.” Beat durations are flexible and should reflect the time each action naturally takes to unfold.
Five test scenes are evaluated side by side: a seated secret agent, a bride running through a church, an elf archer mid-shot, an armored female warrior swinging a sword, and two boxers. Wan 2.2 (with its standard upscaling pass) generally delivers smoother, more polished visuals, while LTX Video running as a raw GGUF output without upscaling shows comparable or occasionally superior motion intensity—particularly for environmental effects like dust and falling leaves. Both models occasionally miss the first beat action. For the boxing scene, neither model fully captures the detail of the reference image.
A recurring practical note: LTX Video requires Kijai’s older VAE version, as the recently uploaded replacement causes colorful static artifacts during sampling. The beat-based prompt structure is confirmed to work with 4-step acceleration LoRAs and is presented as a model-agnostic technique applicable to any image-to-video pipeline.
📺 Source: Veteran AI · Published January 19, 2026
🏷️ Format: Comparison
