Introduction

Video generation used to start with a hopeful one-liner. Now the hard part is the screenplay: actions, camera moves, lighting, and sound that must hold together across multi-shot sequences up to 30 seconds. On Sep 24–25, 2026, Alibaba’s Wan Team (with Nanjing University, USTC, Fudan, and Tsinghua) dropped WanPE — WanPE: Towards Cinematic Prompt Enhancement for Modern Text-to-Video Generation — a 397B-parameter prompt enhancer trained on about 1.05M real-world videos to do that director-level planning for you.

Paper on arXiv:2609.30221, project demos at wan-pe.github.io, and the write-up hit Hugging Face Daily on Sep 25 via taesiri. Lead names include Yubo Zhu, Yawen Shao, Ziyun Dai, and Zixun Fang, with Kai Zhu as project leader.

What WanPE Actually Is

WanPE is not another video generator. It is a cinematic prompt enhancer that expands a natural user request into shot-level plans — actions, camera trajectories, lighting, and sound — sized for modern long-form text-to-video. The flagship is 397B, with reported scales at 4B / 9B / 35B as well.

Treat today’s drop as a research paper + project page with demos. As of this scout, there is no verified public Hugging Face model card shipping downloadable 397B WanPE weights — so this is research/demo territory, not an open-weights weekend pull.

How It Learns Director Taste

Two ideas carry the method:

  1. Video-grounded reverse construction — caption real videos into hierarchical cinematic conditions, then reconstruct the kind of natural requests a user would actually type. Training from reality backward beats inventing “good prompts” from scratch.
  2. Semantic-Consistency GRPO (SC-GRPO) — a reinforcement setup with a 9-dimension reward so the user’s intent survives across shots and over time instead of drifting into pretty-but-wrong scenery.

Ablations: reverse construction clearly beats forward rewriting; SC-GRPO lifts semantic consistency roughly 18.6–23.3 points across scales.

WanPEval and the Numbers That Matter

They curated WanPEval: 249 human-annotated requests and about 11K blind pairwise expert assessments, spanning 5–30s and varied intent.

When WanPE-397B powers Wan3.0’s video generator:

  • vs raw prompts: about +10.66 to +18.84 preference points at 5–15s, and a dramatic +50.86 at 30s
  • leads the evaluated commercial stack at 5–15s (vs Seedance 2.0, MiniMax-H3, Kling 3.0, HappyHorse 1.1, LTX-2.5)
  • stays competitive with Seedance 2.5 at 30s (overall roughly 60.24 vs 59.76)

Cross-generator transfer is the spicy part for creators outside the Wan stack: after format adaptation, WanPE beats native enhancers on LTX-2.5-Base and MiniMax-H3-Base. That H3 transfer is why this matters on ArtRealmAI Gen even before any WanPE weights land in your local folder.

Why Creators Should Care

Long video is where thin prompts fall apart. WanPE’s bet is that the missing layer is cinematic planning in text space — not just prettier adjectives. If your generator already follows complex conditions for 15–30s, a reverse-trained PE that keeps semantics consistent across shots is the difference between “cool clip” and “short that feels directed.”

Caveats stay honest: this is paper + demos today; commercial comparisons are on WanPEval under Wan3.0 (and transfer setups); and open weights for the 397B enhancer are not confirmed public yet. Watch the project page and HF Papers thread for weight announcements.

Original Source

https://arxiv.org/abs/2609.30221

Wan Team / Alibaba Group et al., WanPE: Towards Cinematic Prompt Enhancement for Modern Text-to-Video Generation (arXiv:2609.30221). Project: wan-pe.github.io. HF Papers: huggingface.co/papers/2609.30221 (submitted Sep 24, Daily Sep 25).

Conclusion

WanPE is Alibaba’s answer to the 30-second prompt problem: reverse-train a giant enhancer on a million real videos, then keep user intent alive shot to shot with SC-GRPO. On WanPEval it lifts Wan3.0 hard — especially at 30s — and format-adapted transfer already beats native enhancers on LTX-2.5 and MiniMax-H3. Paper and demos first; weights when (if) they ship.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=wanpe-cinematic-prompt-enhancer

—Aurelia ♡