WanPE: Towards Cinematic Prompt Enhancement for Modern Text-to-Video Generation
WanPE-397B boosts 30-second video-prompt preference by 50.86 points when powering Wan3.0.
WanPE is a 397-billion-parameter prompt-enhancement model trained on 1.05 million real-world videos to produce shot-level cinematic plans for text-to-video generation. It builds plans with video-grounded reverse construction and uses Semantic-Consistency GRPO to keep user intent stable across shots and time. On WanPEval, backed by about 11,000 blind pairwise human assessments, WanPE-397B powering Wan3.0 improves preference over raw prompts by 10.66 to 18.84 points at 5-15 seconds and by 50.86 points at 30 seconds. It leads the evaluated commercial systems at 5-15 seconds and stays competitive with Seedance 2.5 at 30 seconds.
- WanPE is a 397B prompt model trained on 1.05 million real-world videos.
- Semantic-Consistency GRPO preserves user requirements across shots and over time.
- WanPEval uses about 11K blind pairwise assessments spanning 5 to 30 seconds.
- With Wan3.0, preference rises 10.66-18.84 points at 5-15s and 50.86 at 30s.
- WanPE leads tested commercial systems at 5-15 seconds and rivals Seedance 2.5 at 30 seconds.
- AI models
- WanPE-397BWan3.0Seedance 2.5
From the original article
1 min read at huggingface.co
Video generation begins in text space by authoring a cinematic screenplay, then materializes into pixels. As contemporary video generators scale to 30 seconds and faithfully follow complex conditions, the textual prompt largely directs the production, planning how actions, camera trajectories, lighting, and sound unfold across multi-shot sequences.
By Yubo Zhu, Yawen Shao, Ziyun Dai · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.