WanPE: Towards Cinematic Prompt Enhancement for Modern Text-to-Video Generation
Yubo Zhu, Yawen Shao, Ziyun Dai, Zixun Fang, Kai Zhu, Siyang Sun, Haolan Xue, Chuxin Wang and 22 others
Nanjing University · Wan Team, Alibaba Group · University of Science and Technology of China · Fudan University · Tsinghua University
Published September 24, 2026 · on the feed September 25, 2026 · ▲ 37 · arXiv 2609.30221
The abstract
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. In this paper, we present WanPE, a 397B-parameter prompt enhancement model trained on 1.05M real-world videos to master director-level cinematic planning. WanPE formulates shot-level cinematic plans via video-grounded reverse construction and employs Semantic-Consistency GRPO (SC-GRPO) to faithfully preserve user requirements across shots and over time. To benchmark this capability, we curate WanPEval, a human-annotated testbed covering durations from 5 to 30 seconds across varying intent granularities, supported by approximately 11K blind pairwise assessments. When powering Wan3.0's video generator, WanPE-397B boosts human preference over raw user prompts by 10.66-18.84 points at 5-15 seconds and by a dramatic 50.86 points in the 30-second arena. Ablation studies show that reverse construction demonstrates clear superiority over forward rewriting, while SC-GRPO robustly preserves semantic fidelity across model scales. Ultimately, WanPE leads all evaluated commercial offerings at 5-15 seconds and remains competitive with Seedance 2.5 at 30 seconds.
The Brief
2:27A short podcast summary of the paper. Written and voiced by AI.
Transcript
WanPE. Towards Cinematic Prompt Enhancement for Modern Text-to-Video Generation. By Yubo Zhu and twenty-nine others, from Nanjing University, the Wan Team at Alibaba Group, and three other institutions. Published on the twenty-fourth of September. No code repository is linked to it.
The authors picture video generation as beginning in text, with a cinematic screenplay written first and only then turned into pixels. As video generators stretch to thirty seconds and follow complex conditions faithfully, that text carries more of the weight. The prompt now largely directs the production. It plans how actions, camera trajectories, lighting, and sound unfold across a sequence of several shots. A short user request is a long way from a plan like that.
WanPE is a prompt enhancement model built to close that distance, to master planning at the level of a director. It has three hundred and ninety-seven billion parameters and is trained on one point zero five million real-world videos. Rather than rewriting a prompt forward, it formulates a cinematic plan shot by shot, through what the authors call video-grounded reverse construction.
Keeping the user's intent intact is the second concern. For that, WanPE uses Semantic-Consistency GRPO, or SC-GRPO, to preserve user requirements faithfully across shots and over time. To measure the capability, the authors also curate WanPEval, a human-annotated testbed. It covers durations from five to thirty seconds and several levels of intent granularity, and it is backed by approximately eleven thousand blind pairwise assessments.
Paired with the video generator of Wan3.0, WanPE-397B raises human preference over raw user prompts by between ten point six six and eighteen point eight four points, for videos of five to fifteen seconds. In the thirty-second arena the gain is fifty point eight six points.
Ablations back both components. Reverse construction shows clear superiority over forward rewriting, and SC-GRPO preserves semantic fidelity robustly across model scales. Against commercial offerings, WanPE leads all of those evaluated at five to fifteen seconds, and remains competitive with Seedance 2.5 at thirty seconds.
Here the prompt is treated as the screenplay rather than a caption. Grounding the plan in real videos by reverse construction, then holding it to what the user asked for, is the pairing the ablations support. The largest margin arrives at the longest duration, where a plan has the most shots to hold together.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
The prompt now directs the whole film. Writing it well is the problem.
You are building the text stage of a video generator that makes clips up to 30 seconds long. A short user request must become a shot-by-shot script: actions, camera, lighting and sound.
The road ahead
- 1 · The problem ~7 min A short request must grow into a full shot-by-shot script, and the script's style has to match the text the video generator learned from.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Build training pairs backward from real videos, so the targets look like the generator's captions, then use RL to keep every user requirement.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Collect real clips, caption them, derive requests, fine-tune on the pairs, then refine with SC-GRPO.
- 6 · The experiment ~6 min You’ll guess: In the 30-second arena, by how many points did WanPE-397B raise human preference over raw user prompts?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.