paper web signal

WanPE: 397B Cinematic Prompt Model Delivers 50-Point Human-Preference Gain on 30-Second Video Generation

Summary

Demonstrates that a dedicated 397B prompt-enhancement stage—translating bare user intent into shot-by-shot cinematic plans—yields outsized quality gains on long-form video generation, competitive with commercial leaders, suggesting the bottleneck in video AI is increasingly in the text rather than the pixels.

Shared on Bluesky by 1 AI expert