AIAny
Icon for item

WanPE: Towards Cinematic Prompt Enhancement for Modern Text-to-Video Generation

Transforms user prompts into shot-level cinematic directions for text-to-video generation, using a 397B prompt-enhancer trained on 1.05M videos and SC-GRPO to preserve semantic consistency across shots; evaluated on WanPEval (5–30s) with large human-preference gains.

Introduction

Most text-to-video systems treat prompts as short, descriptive seeds — but as generators extend to multi-shot, cinematic outputs, prompts must plan camera, staging, lighting and temporal continuity across shots. WanPE reframes prompt enhancement as director-level planning rather than simple rewriting, and quantifies how much better video quality and intent fidelity benefit from that shift.

Key Findings
  • Large-scale reverse-construction training: WanPE uses a reverse-supervision pipeline to derive hierarchical, shot-level cinematic conditions from high-quality videos and reconstruct compatible user requests, enabling planner-style outputs rather than surface rewrites.
  • Scale and data: The paper reports a 397B-parameter prompt enhancement model trained on ~1.05M real-world videos, demonstrating consistent gains as model size grows (4B→397B ablations included).
  • Semantic-Consistency GRPO (SC-GRPO): A nine-dimensional reward optimized with group relative policy optimization to penalize omissions, incorrect bindings between subjects/actions/dialogue, and temporal inconsistencies across shots, improving requirement preservation.
  • Evaluation with WanPEval: A human-annotated 5–30s testbed and ~11K pairwise assessments show WanPE-397B boosts human preference over raw prompts by ~10.66–18.84 points at 5–15s and by ~50.86 points at 30s; reverse construction outperforms forward rewriting and SC-GRPO yields notable semantic-fidelity gains.
  • Transferability: WanPE formats and transfers across downstream video generators after adaptation, and leads or competes with contemporary systems on short-clip benchmarks.
Who It's For and Tradeoffs
  • Great fit if you develop or evaluate text-to-video pipelines and need structured, multi-shot prompt conditioning that enforces temporal and semantic fidelity across a short cinematic horizon.
  • Look elsewhere if you only require single-frame or very short (1–3s) generation where lightweight prompt templates suffice, or if computational/latency budgets prohibit large prompt-enhancer models at inference time.
Method Snapshot
  • Reverse construction: extract hierarchical cinematic conditions from videos, then learn to map user intents to those conditions, encouraging planner-like outputs.
  • SC-GRPO reward: combines omission/alteration penalties and binding/temporal consistency checks into group-relative policy optimization to keep enhanced prompts faithful to user intent.

These elements position WanPE as a focused study on elevating prompt engineering from enrichment to explicit cinematic planning for modern T2V systems.

Information

  • Websitearxiv.org
  • AuthorsYubo Zhu, Yawen Shao, Ziyun Dai, Zixun Fang, Kai Zhu, Siyang Sun, Haolan Xue, Chuxin Wang, Tingyu Weng, Jingming Luo …
  • Published date2026/09/24

Categories

More Items

Provides WROP: a 1.5M-sample synthetic video corpus and a 300-question exam for training and evaluating object permanence and solidity in video world models. Includes 150 Blender task generators, a human Elo benchmark across 14 models, and a fine-tuned 16B continuation model (PWM-WROP).

Introduces a rubric-based approach for video reward modeling that generates explicit, query-adaptive evaluation criteria before scoring to reduce scalar drift; proposes RGPO, a two-stage training (seed warm-up + joint optimization) that achieves strong pointwise and pairwise evaluation with high data efficiency.

Surveys memory mechanisms for autoregressive video generation, framing memory as persistent historical information that influences future generation. Organizes work by Forms, Functions, Operations, Learning, and Evaluation, and synthesizes challenges for long-horizon consistency and memory-aware learning.