Video generation starts in text space, where a cinematic screenplay is authored before it becomes pixels. As modern video generators scale to 30 seconds and follow increasingly complex conditions, the textual prompt increasingly drives production — planning actions, camera trajectories, lighting, and sound across multi-shot sequences.
A new paper on Hugging Face presents WanPE, a 397-billion-parameter prompt enhancement model trained on 1.05 million real-world videos to perform director-level cinematic planning. The model formulates shot-level cinematic plans through video-grounded reverse construction and applies Semantic-Consistency GRPO (SC-GRPO) to preserve user requirements faithfully across shots and over time.
To measure this capability, the authors curated WanPEval, a human-annotated testbed spanning durations from 5 to 30 seconds across varying intent granularities, supported by roughly 11,000 blind pairwise assessments.
When paired with Wan3.0's video generator, WanPE-397B reportedly raises human preference over raw user prompts by 10.66–18.84 points at 5–15 seconds and by 50.86 points in the 30-second arena. Ablations indicate reverse construction outperforms forward rewriting, while SC-GRPO preserves semantic fidelity across model scales.
The paper states WanPE leads all evaluated commercial offerings at 5–15 seconds and remains competitive with Seedance 2.5 at 30 seconds.