WanPE: Advancing Cinematic Prompt Refinement for Contemporary Text-to-Video Generation

Meta description: WanPE is a 397B-parameter prompt enhancement model trained on 1.05M real-world videos for cinematic planning and 30-second text-to-video generation benchmarks.

Tags: text-to-video, prompt engineering, WanPE, video generation, cinematic planning

Featured image: Abstract friendly illustration of a text prompt expanding into storyboard frames and moving video clips, with no real people, no logos, and no text.

For creators, developers, and researchers working on text-to-video tools, the paper's authors present WanPE, a 397B-parameter prompt enhancement model trained on 1.05M real-world videos to turn short user instructions into shot-level cinematic plans. The source does not state a publication date. The core idea is simple: video generation starts in text and ends in pixels, and WanPE writes a director-like plan before the generator renders those pixels.

Why Prompts Decide Video Quality

As text-to-video systems have grown from single clips toward 30-second sequences, the prompt has become more than a description. It is a production order. A user may ask for a rainy street scene, a character reaction, a camera push-in, and a shift from dusk to night, but the generator still needs to know which shot comes first, how long each shot lasts, and how continuity should be preserved. Contemporary generators can follow complex conditions, yet raw prompts often leave these choices implicit.

That gap matters because longer videos amplify small planning mistakes. If a model misunderstands the order of actions, a shot may feel disconnected. If it overwrites a requested detail with a more cinematic alternative, the output may look polished but no longer match the user's intent. WanPE is positioned as a middle layer between the user and the video generator: it expands a brief request into a structured plan, then the generator materializes that plan as pixels.

How WanPE Builds Shot Plans

WanPE does not merely rewrite a sentence into a longer sentence. The paper frames the task as director-level cinematic planning and describes video-grounded reverse construction: the model learns from real-world videos by working backward from finished footage to the kind of shot plan that could have produced it. This gives it a practical sense of cinematic structure instead of relying only on written examples. The output is a set of shot-level plans that specify what happens in each shot and how shots connect over time.

The second key piece is Semantic-Consistency GRPO, abbreviated SC-GRPO. The paper uses it to keep the user's original requirements stable across shots and throughout the sequence. This matters because a prompt enhancement model can otherwise drift: it may add attractive cinematic details while losing a required character, action, or mood. SC-GRPO is presented as a consistency mechanism, and the paper's ablation results say it robustly preserves semantic fidelity across model scales.

What The Benchmark Results Show

The evaluation uses WanPEval, a human-annotated test bed covering durations from 5 to 30 seconds across varying intent granularities. The dataset is supported by approximately 11K blind pairwise assessments. In a blind pairwise setup, raters compare outputs and choose a preference without knowing which system generated each result, which can reduce bias from brand labels.

The paper reports that when WanPE-397B powers Wan 3.0's video generator, human preference improves over raw user prompts by 10.66-18.84 points at 5-15 seconds. The largest reported gain appears in the 30-second arena, where the preference increase is 50.86 points. The benchmark also compares against commercial offerings. At 5-15 seconds, WanPE leads all evaluated commercial offerings in the paper's setup. At 30 seconds, it remains competitive with Seedance 2.5.

Another ablation result shows reverse construction demonstrates clear superiority over forward rewriting. The practical meaning is that learning from video evidence appears more effective than simply rewriting the prompt in text form. The paper does not provide a public release date, pricing, API access, or details about which products will ship the model.

Who should use it / who should skip it

This paper is most useful for researchers and AI video practitioners evaluating how prompt planning affects long sequences. It gives concrete evidence that a 397B-parameter enhancement layer trained on 1.05M real-world videos can improve human preference, especially at 30 seconds, and that video-grounded reverse construction may be more effective than forward rewriting.

It is less useful for readers looking for an immediate consumer product. The source does not state when the model was released or whether it will be available through an API or commercial platform. Until those questions are answered, WanPE is best read as a technical signal: the next competition in text-to-video may be won by the model that can plan a shot sequence faithfully, not only render a single frame well.