Alibaba's Wan video generation team (@Alibaba_Qwen) has released WanPE (arXiv: 2609.30221, Project: wan-pe.github.io), a 397B-parameter prompt enhancement architecture powering the upcoming Wan3.0 video foundation model. Addressing narrative incoherence and camera drift when scaling text-to-video generation to 30-second sequences, WanPE leverages a video-grounded reverse construction pipeline and Semantic-Consistency GRPO (SC-GRPO) to generate multi-shot cinematic screenplays without deviating from user intent, driving a 50.86-point surge in human preference ratings over raw prompts.

Key Takeaways

  • ✓Releases WanPE-397B, the largest open cinematic prompt enhancement model trained on 1.05M real-world cinematic videos.
  • ✓Introduces Semantic-Consistency GRPO (SC-GRPO), eliminating semantic drift where RL rewrites unintentionally alter core user constraints.
  • ✓Employs reverse shot-level construction to plan camera trajectories, lighting, and multi-shot transitions across 5-30s videos.
  • ✓Boosts human preference scores by 50.86 points in 30-second video generations when paired with Alibaba Wan3.0.
  • ✓WanPEval benchmark dataset comprising 11,000 blind pairwise human assessments published on GitHub and project page.
🔗

Project Links & Resources

Direct Access
Direct access to official project resources and documentation
📝
WanPE 官方项目主页与视频生成演示
wan-pe.github.io/
↗
🌐
ArXiv 论文预印本 (2609.30221)
arxiv.org/abs/2609.30221
↗
📝
Hugging Face Daily Papers 讨论主页
huggingface.co/papers/2609.30221
↗
🔬

In-Depth Technical Analysis

Core Background & Industry Pain Points As generative video models scale toward sustained 30-second sequences, video synthesis increasingly demands director-level temporal planning. Naive user prompts (e.g., 'a detective walking in a rainy alley') lack kinematic instructions, causing models to generate monotonous pans or severe temporal distortions. Conversely, naive LLM prompt rewrites suffer from hallucinated embellishments, discarding user-specified subjects or color constraints. ### Architecture Highlights & Internals Alibaba presents WanPE, a 397B-parameter prompt enhancement engine trained on 1.05M curated cinematic sequences. Through video-grounded reverse construction, WanPE learns to reverse-engineer dense director screenplays—prescribing shot transitions, orbital camera rigs, key lighting, and Foley sound cues. Crucially, the authors deploy Semantic-Consistency GRPO (SC-GRPO), anchoring policy optimization to strict semantic mutual information bounds to prevent prompt drift while maximizing cinematographic visual richness. ### Authoritative Benchmarks & Measured Scores Evaluated on WanPEval across 11,000 blind pairwise human assessments: WanPE-397B lifts visual preference over raw prompts by 10.66-18.84 points on 5-15s clips, and by 50.86 points in 30s sequences when powering Wan3.0. Across camera choreography and temporal coherence, WanPE-397B leads all evaluated commercial rewrite APIs while matching Seedance 2.5 on long-horizon outputs. ### Developer Hands-on Guide Creators and video pipeline architects can inspect comparative video reels at https://wan-pe.github.io/ and review the SC-GRPO prompt formatting schemas on arXiv and Hugging Face Papers.