WanPE is a 397-billion parameter prompt enhancement model trained on 1.05 million real-world videos to master director-level cinematic planning. It uses a reverse construction method: it first derives a hierarchical cinematic condition from each video, then reconstructs a compatible user request. It optimizes prompt generation with Semantic-Consistency GRPO (SC-GRPO), a reinforcement learning method that uses a nine-dimensional reward to penalize omissions, alterations, and temporal inconsistencies.
When paired with Wan3.0's video generator, WanPE improved human preference over raw user prompts by 10.66 to 18.84 percentage points for 5- to 15-second videos, and by 50.86 points for 30-second videos. It outperformed all evaluated commercial offerings at 5-15 seconds and remained competitive with Seedance 2.5 at 30 seconds.
This work shifts prompt enhancement from simple description to full cinematic planning, including camera, lighting, and multi-shot sequences. As video generators handle longer contexts, such structured prompting becomes crucial. The reverse construction approach ensures the enhancer's output distribution matches that of real video captions, addressing a key mismatch in prior methods.