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

1 Introduction

• Prompt enhancement shifts from descriptive enrichment to cinematic direction, planning actions, camera, lighting, dialogue, and sound across shots and time.The prompt becomes a production blueprint as generators handle longer contexts and more complex instructions.Source: §1 · pp. 1–2
• Forward rewriting creates a distribution mismatch with video-grounded generator training, while cinematic plans require global preservation of user requirements across shots and time.Requirements can be omitted, altered, bound to the wrong subject or shot, or contradicted later.Source: §1 · p. 2
• Up to 18.84 points at 5-15 seconds and 50.86 points at 30 seconds: WanPE-397B improves over raw prompts.It leads evaluated commercial offerings at 5-15 seconds and remains competitive with Seedance 2.5 at 30 seconds.Source: §1 · p. 2
• Reverse construction outperforms forward rewriting, SC-GRPO improves semantic consistency by 18.6-23.3 points across scales, and WanPE transfers across generators after format adaptation.The reported semantic-consistency improvement holds across model scales.Source: §1 · pp. 2–3
• WanPE learns cinematic structure from real-world videos through reverse construction and uses SC-GRPO to preserve requirements throughout the plan.Its architecture treats text as a blueprint coordinating actions, camera choreography, audio, and narrative across shots.Source: §1 · pp. 2–3
• WanPEval covers 5-30-second requests with varying intent granularities and evaluates 4B-397B variants using approximately 11K blind pairwise video assessments.The testbed is introduced to assess different request complexities and generation durations.Source: §1 · pp. 2–3

2.1 Problem Overview and Formulation

• The ideal enhancer target restricts the video-grounded caption distribution to conditions that satisfy all user-request constraints.The enhancer is trained to match this target over the user-prompt distribution.Source: §2.1 · p. 3

2.2 Video-Grounded Supervised Fine-Tuning

• Reverse construction first captions real-world videos, then reconstructs compatible user requests to train cinematic conditions closer to video-grounded captions.The curated collection contains approximately 1.05M clips; captions organize video summaries and timestamped shot descriptions.Source: §2.2 · p. 4

2.3 Semantic-Consistency GRPO

• SC-GRPO penalizes omissions, alterations, incorrect subject-action-dialogue bindings, and temporal inconsistencies across shots.Its semantic reward assesses nine dimensions, including subjects, actions, dialogue, sound, camera, lighting, spatial relations, and scene.Source: §2.3 · p. 5

2.4 WanPEval: Testbed and Evaluation

• WanPEval spans 5–30-second generations and request granularities from concise high-level intents to detailed shot-level instructions.Its curated requests are human-written, reviewed for clarity and feasibility, deduplicated, and screened for training-data overlap.Source: §2.4 · pp. 5–6
• Evaluation combines text-level semantic-consistency scoring with anonymous pairwise video preferences from 60 experts across film-related disciplines.The preference protocol randomizes method identities and presentation order, treats both-good and both-poor judgments as ties, and reports Bradley–Terry scores against a mean-strength opponent.Source: §2.4 · p. 6

3.1 Overall Comparison

• Under the same Wan3.0 video generator, WanPE-397B improves over the original-request baseline by 10.66, 16.28, and 18.84 points at 5, 10, and 15 seconds, then raises S from 9.38 to 60.24 on the 30-second subset.Source: §3.1 · p. 7
• WanPE-397B ranks first on intent-, scene-, and shot-level requests, with scores of 52.44, 49.26, and 49.32.It improves on no prompt enhancement by 18.55, 16.97, and 13.45 points, respectively.Source: Figure 3 · p. 7, §3.1 · p. 7

3.2 Reverse-Constructed Supervision vs. Forward Prompt Expansion

• 49.86 overall S for WanPE-397B-SFT exceeds Forward Rewriting by 10.37 points and Forward-target SFT by 14.69 points.The model also ranks first across all seven content categories.Source: Table 2 · p. 8, §3.2 · p. 8

3.3 Transferability of Cinematic Planning Across Video Generators

• After format adaptation, WanPE-397B improves S over native enhancers by 14.45 points on LTX-2.5-Base and 5.17 on MiniMax-H3-Base.The experiments convert its outputs to each generator’s native format while retaining their cinematic content.Source: Table 2 · p. 8, §3.3 · p. 8

3.4 Semantic Consistency and SC-GRPO Analysis

• 18.6–23.3 points of semantic-consistency improvement occur with SC-GRPO across four model sizes, with gains at every evaluated duration.Mean training reward rises across the four scales, and larger models reach higher final values.Source: §3.4 · p. 8, Figure 4 · p. 8, Table 3 · p. 9
• SC-GRPO raises WanPE-397B’s overall preference score from 42.70 to 49.69 and improves results across all seven content categories.The comparison uses the same Wan3.0 generator and differs only in the enhanced conditions.Source: Table 3 · p. 9, §3.4 · p. 9

Text-to-video generation

• Text-to-video systems have progressed from early latent video models to large diffusion transformers, alongside closed-source systems such as Sora 2, Kling, Seedance 2.5, and Wan3.0.Source: §4 · p. 9

Prompt enhancement for text-to-video generation

• T2V prompt enhancement includes learned rewriting and refinement [3–12], preference or supervised optimization [4] [5], retrieval-based alignment [34], and inference-time critique [10] [12].The cited methods also explore multimodal training, semantic verification, and physics-aware prompting.Source: §4 · p. 9

Proprietary prompt enhancers

• Commercial systems increasingly rely on native prompt enhancers, but most designs and training recipes remain largely undisclosed.MiniMax-H3’s H3-Context-IR is separately accessible, while other listed enhancers are integrated into end-to-end APIs.Source: §4 · p. 9

• WanPE treats prompts as blueprints orchestrating actions, camera choreography, audio, and narrative across shots and time before pixel rendering.The conclusion reports that scaling from 4B to 397B consistently improves semantic fidelity and downstream video quality; reverse construction, SC-GRPO, and format adaptation also show supported gains.Source: §5 · p. 10

A.2 Shot- and Sequence-Level Organization

• Shot-level clips capture actions, composition, camera work, lighting, and sound, while sequences expose how these elements develop through transitions and narrative progression.The two forms provide complementary supervision for fine-grained realization and long-range structure.Source: §A.1 · p. 14, §A.2 · p. 14

A.1 Collection Scope

• The collection includes publicly available or licensed videos, divided into clips of no more than 30 seconds for structured cinematic captioning.Source: §A.1 · p. 14

A.3 Quality Filtering and Curation

• Curation first filters technical validity, visual quality, and motion quality, then assesses shot completeness and performance or sequence coherence and cinematic organization.Sequence-specific checks include visual-style consistency for animation and effective information delivery in motion graphics.Source: §A.3 · p. 14

A.4 Resulting Corpus

• Approximately 1.05M clips of up to 30 seconds across ten content dimensions support cinematic conditions at local and long-range temporal scales.The corpus combines fine-grained shot-level evidence with realized multi-shot organization.Source: §A.4 · p. 14

B Video-Grounded Reverse Construction

• Video-grounded reverse construction converts curated videos into cinematic targets, with the following subsections detailing caption structure, adaptation, and quality control.The retained captions become the targets for supervised fine-tuning.Source: B Video-Grounded Reverse Construction · p. 14, §B.3 · p. 15

B.1 Hierarchical Caption Structure

• Each target combines a video-level summary with timestamped, ordered shot descriptions covering visual, narrative, camera, and audio elements.Summaries capture overall content, setting, style, perspective, pacing, and audio design; shot descriptions cover composition, subjects, actions, lighting, camera movement, transitions, dialogue, music, and sound effects.Source: §B.1 · p. 15

B.2 Category-Adaptive Captioning

• Caption schemes adapt to content, emphasizing motion and audio for large-motion videos, synchronization for music, contextual transcription for dialect, and text layout for rich-text videos.General videos prioritize comprehensive scene coverage; animation, advertisements, visual effects, and complex camera movement receive tailored descriptive emphasis.Source: §B.2 · p. 15

B.3 Caption Quality Control

• Captions are checked for structural completeness, valid timestamps, and consistency with the source video; failures trigger regeneration before captions are retained as fine-tuning targets.Distinct system prompts and preprocessing tags reinforce relevant descriptive dimensions.Source: §B.3 · p. 15, §B.2 · p. 15

C WanPEval Design and Construction

• WanPEval contains 249 curated requests spanning diverse content categories, target durations, aspect ratios, and request granularity.The following subsection presents the benchmark distribution and representative intent-, scene-, and shot-level requests.Source: C WanPEval Design and Construction · p. 15

C.1 Benchmark Coverage and Distribution

• Figure 5 summarizes 249 requests across seven content categories, four target durations from 5 to 30 seconds, two aspect ratios, and three request-granularity levels.The distribution also characterizes shot-count specifications, audio components, and cinematic language.Source: §C.1 · p. 15, Figure 5 · p. 16

C.2 Representative Requests Across Granularity Levels

• Intent-level requests give a high-level scenario, scene-level requests describe event progression, and shot-level requests specify shots and elements such as camera movement and dialogue.The examples illustrate variation in request detail across the three levels.Source: §C.2 · p. 17

D Additional Quantitative Results

• The additional results report baseline-specific preferences and exact duration and request-granularity breakdowns for the 5–15-second WanPEval subset.They supplement Table 1 and Figure 3.Source: D Additional Quantitative Results · p. 17

D.1 Baseline-Specific Preference Results

• Table 4 reports preference scores across seven content categories and overall, comparing each external system with the original-request baseline and WanPE-397B.Scores aggregate valid pairwise judgments within each block, and the gray row indicates the downstream generator used by the latter two systems.Source: §D.1 · p. 17, Table 4 · p. 18

D.2 Fine-Grained Preference Breakdowns

• WanPE-397B attains the highest central preference score at every duration and request granularity.The corresponding intervals are listed in Tables 5 and 6 in Slower/S/Supper order.Source: §D.2 · p. 18
• Table 5 lists exact preference intervals by duration, with entries ordered Slower/S/Supper.These values are represented by the markers and intervals in Figure 3’s left panel.Source: Table 5 · p. 18, §D.2 · p. 18
• Table 6 lists exact preference intervals by request granularity, with entries ordered Slower/S/Supper.These values are represented by the markers and intervals in Figure 3’s left panel.Source: Table 6 · p. 18, §D.2 · p. 18