acceptodds
Under review as a conference paper at ICLR 2027

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

10 views

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 demo is available here.

Primary area:
generative models
Keywords:
prompt enhance model, text to video generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_VFWSOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces WanPE, a prompt enhancement model that transforms user requests into detailed, temporally structured prompts for text-to-video generation. It constructs SFT pairs by reverse-generating user requests from captions of 1.05 million real-world videos, then applies semantic-consistency GRPO to improve requirement preservation. Experiments with 4B–397B models on WanPEval, covering 5–30-second videos, report improvements in semantic fidelity and downstream video quality.

Critical strengths
  • The paper studies model scaling from 4B to 397B, showing generally stronger semantic fidelity and downstream video performance at larger scales.

  • The paper constructs a large-scale, video-grounded SFT dataset from approximately 1.05 million real-world clips, providing supervision for diverse cinematic content and both single-shot and multi-shot structures.

Critical weaknesses
  1. The method novelty appears limited, as the approach largely combines existing data construction techniques, supervised fine-tuning, and GRPO. The paper should more clearly distinguish its contributions from closely related prompt enhancement methods. Moreover, this paper does not conduct a system and qualitative comparison with this method and existing PE methods.

  2. The paper claims “director-level cinematic planning,” but evaluates it only indirectly through semantic fidelity and overall video preference. Separate assessments of shot organization, pacing, and audiovisual coordination are needed to establish better planning than strong rewriting baselines.

  3. The paper does not analyze whether the reverse-generated requests resemble real user requests in distribution, including token length and category coverage. Also, the training requests and PE text are insufficiently analyzed, with no statistics on token-length distributions, expansion ratios, category and vocabulary distributions, or diversity in phrasing.

  4. No data-scaling experiments are reported. Consequently, it is unclear how much of the dataset is necessary, whether performance saturates with substantially less data, or whether reverse construction offers better data efficiency than forward construction.

  5. Although the paper includes a Gemini-based direct rewriting baseline, its inference configuration is not sufficiently documented. The complete system prompt, reasoning settings, and decoding parameters are needed to assess and reproduce this comparison.

  6. The demonstrations primarily compare generation with and without WanPE. It is no wonder for me that w/o PE performs worse than w/ PE, so the demonstration is weak. Paired examples showing both the rewritten text and generated videos from direct LLM rewriting (or, Forward-target SFT) versus WanPE would make the method-specific improvements easier to inspect.

  7. The training ablations are incomplete: the paper does not report forward-target SFT followed by the same SC-GRPO procedure. This would help determine whether the advantage of reverse construction persists after both models receive equivalent semantic-consistency training.

  8. The two semantic evaluators (Gemini and GPT) disagree substantially about the final model’s reliability: the reported perfect-score rates are 85.5% and 7.2%. Although both evaluators support improvements from RL, this discrepancy requires explanation and human adjudication.

  9. The human evaluation lacks important statistical and annotation details, especially confidence intervals and the number of judgments per pair.

Questions for authors
  1. Why not use Qwen as Direct rewriting Baseline?
  2. Does the SFT training data contain CoT?
  3. Will this paper release the SFT training data and RL prompts?
5 comments
AuthorsOpenReview ↗

Response to Reviewer VFWS (Part 1/5)

We thank the reviewer VFWS for the careful and constructive comments. We will try our best to answer all your questions. Please let us know if you still have further concerns, or if you are not satisfied with the current responses, so that we can further update the response ASAP.


W1: method novelty and qualitative comparison

A: We thank the reviewer VFWS for raising this concern for these.

To the best of our knowledge, we are the first to openly investigate how to train native prompt enhancers in modern commercial video generation systems. These systems incorporate a native PE and a video generator. When the official system APIs are called, user requests first pass through the native PE and then to the video generator. However, the training methods and details of such native PEs remain largely undisclosed. We identify two core challenges in native PE design for modern DiTs, and address them through video-grounded reverse SFT and SC-GRPO. More broadly, as video DiTs continue to advance, we believe that the PE should evolve from an optional auxiliary module into an indispensable director for fine-grained, holistic planning.

Reviewer qcA3 recognized that “the video-grounded reverse SFT design is novel and well motivated,” while Reviewer DjxV described it as “an interesting alternative for learning video-grounded cinematic planning.”


System and qualitative comparison.

We compare along two axes: (i) native PEs of modern commercial video generation systems, and (ii) previous general-purpose PE methods.

(i) Native PEs of modern commercial systems.

  • Compare with the PE deployed in the production version of Wan3.0. To further validate the effectiveness of our native PE training approach, we report a previously completed internal evaluation. We invited 100 participants, including short-video creators, short-drama professionals, and general users, to test our system. We randomly sampled 2,400 prompts from their test inputs and compared WanPE with the native PE deployed in the production version of Wan3.0 (2026-08-27), following the human blind-evaluation protocol in Section 2.4. We compared two settings: (i) enhancing user prompts with a locally deployed WanPE-397B and submitting them to the Wan3.0 official API with the production PE disabled; and (ii) submitting the original user prompts to the same API with the production PE enabled. WanPE performed comparably to the PE deployed in production, with preference scores SS of 51.29% and 50.02%, respectively. This comparison demonstrates the effectiveness of our approach for training native prompt enhancers.

  • Compare with the native PEs of MiniMax-H3 and LTX-2.5. In Section 3.3, we have compared WanPE against H3-Context-IR and LTX-2.5-PE on the same video generators, H3-Base and LTX-2.5-Base, respectively. The results in Table 2 demonstrate WanPE’s strong transferability.

  • For other commercial systems, the native PEs are not independently accessible.

(ii) Previous General-purpose PE methods

In the initial version, we did not compare against excellent prior PE methods such as RAPO, VPO, Prompt-A-Video and SCMAPR because we considered such a comparison unfair to them:

  • Different design targets. Prior PE methods were developed for previous-generation video models that natively supported only short-duration generation. Given their limited context capacity and relatively weak instruction-following capabilities of video models, prompt enhancement primarily served to enrich visual details. Their ability to generalize across diverse user inputs may also be limited.

Following the reviewer’s suggestion, we compare baselines on Wan3.0’s video generator with the 30-second subset of WanPEval, and on H3-Base with the 15-second subset due to its maximum generation duration, following the blind evaluation protocol in Section 2.4. Each baseline is compared with WanPE-9B (parameter is comparable to those of the trainable baseline PEs) in a separate 1v1 battle, and baselines are not compared against each other. The preference scores SS are as follows:

MethodActionAnim.SpeechAd.Sing & DanceDramaKnow.Overall
downstream generator: Wan3.0’s video generator
+RAPO00000000
+VPO000053.8501.64
+Prompt-A-Video07.140007.6902.46
+SCMAPR05.565.560153.8504.55
+WanPE-9B78.269093.5910077.5076.9292.8684.54
downstream generator: MiniMax-H3-Base
+RAPO01.727.14004.5501.72
+VPO8.333.4508.3312.509.092.505.17
+Prompt-A-Video06.9008.336.254.552.504.60
+SCMAPR8.3312.0714.298.3312.504.5559.20
+WanPE-9B79.1775.8673.2193.7582.8167.058077.59

The results show that WanPE-9B leads across all categories. We have also updated the demo page in the anonymous link, which directly visualize the comparison with prior methods.

AuthorsOpenReview ↗

Response to Reviewer VFWS (Part 2/5)

W2: Separate assessments of shot organization, pacing, and audiovisual coordination

A: We thank the reviewer VFWS for this suggestion!

Following the reviewer's suggestion, we report separate assessments on these aspects using PEBench [1], a third-party benchmark that directly evaluates enhanced prompts. We report all nine Directorial Creation (DC) criteria and their aggregate. Shot organization corresponds to Storyboard and Timing Completeness (STC), pacing to Editing and Transition Design (ETD), and audiovisual coordination to Sound Design Integration (SDI). The remaining criteria are Camera Design Specificity (CDS), Lighting and Color Specification (LCS), Narrative Development and Engagement (NDE), Core-Intent Focus (CIF), Subject and Performance Specificity (SPS), and Visual Style Specification (VSS). The results for T2V PE tasks are as follows:

MethodSTC ↑CDS ↑ETD ↑LCS ↑NDE ↑CIF ↑SDI ↑SPS ↑VSS ↑Overall ↑
RAPO0.001.401.802.401.4014.401.003.604.200.93
Prompt-A-Video-OS1.0019.0014.4030.8022.4086.809.6037.4054.4018.99
Prompt-A-Video-Cog0.6023.0021.0058.2048.4083.8014.6065.2071.2029.66
VPO0.4025.2027.4064.2048.8089.606.6075.2070.0033.07
SCMaPR6.2055.2041.6048.6061.0099.2013.2074.8070.6049.18
LingBot-Video-PE40.6062.8046.2035.6060.2065.804.6077.8070.4048.66
LTX-2.5-PE2.4076.6036.4066.8065.4094.2061.2082.6077.4057.69
H3-Context-IR12.2089.0055.4076.8076.4094.0093.6086.4073.6076.73
WanPE-35B59.8080.2063.6075.6066.6094.2083.0080.4078.6078.15
WanPE-397B66.6081.0063.2080.2068.6091.2077.6081.6082.2078.97

The results show that WanPE-397B attains the highest overall score. Among them, H3-Context-IR is the native PE of the commercial system MiniMax-H3, accessible only through its API, and its implementation and training details are not disclosed. We also compared WanPE with H3-Context-IR on H3-Base as the same downstream generator, where WanPE achieved higher human preference (Table 2(iii)).

[1] Shao, Yawen, et al. "Towards Unified Evaluation of Prompt Enhancers for Video Generation." arXiv preprint arXiv:2610.11736 (2026).

AuthorsOpenReview ↗

Response to Reviewer VFWS (Part 3/5)

W3: The paper does not analyze whether the reverse-generated requests resemble real user requests in distribution.

A: We thank the reviewer VFWS for raising this concern.

We believe that generalization across diverse categories of user inputs is a primary consideration when developing native PEs for modern commercial video generation systems. Therefore, we include stylistically diverse data in both data construction and evaluation, as detailed below.

training data construction.

SFT data. Video-grounded targets yy are derived from real-world videos spanning ten content categories, each captioned with a category-specific instruction (Appendix B). The few-shot demonstrations used for reconstruction are drawn from a 2K human-written request pool stratified by category, covering granularities from one-sentence intents to detailed shot-level instructions. We report statistics of yy and xx over the full SFT set below.

  • Statistics of video-grounded targets yy:

    • Category and language: ten content categories; captions are written in Chinese (50.8%) or English (49.2%), and each xx is reconstructed in the language of its caption.
    • Length: min 275 tokens, mean 1,683, median 1,515, max 6,809.
    • Duration: median 14.0 s; ≤5 s 3.4%, 5–10 s 12.5%, 10–15 s 37.5%, 15–20 s 21.6%, 20–30 s 25.0%.
    • Speech: 60.3% of targets contain speech transcribed verbatim from the source audio, covering 25 languages: en 40.6%, zh 35.7%, ja 4.5%, ru 2.9%, es 2.1%, tr 1.8%, ko 1.7%, fr 1.5%, de 1.4%, tl 1.0%, it 1.0%, ar 0.8%, pt 0.7%, id 0.5%, nl 0.5%, vi 0.5%, th 0.5%, pl 0.4%, ms 0.4%, fi 0.4%, cs 0.3%, sv 0.3%, da 0.3%, hu 0.2%, ro 0.2%.
  • Statistics of reconstructed requests xx:

    • The four tables below show that xx spans a broad range of lengths and constraint types, with diversity comparable to human-written prompts.
Token lengthmeanmedianmax≤10≤50≤200≤500≤1,000>1,000
xx223931,8172.5%25.8%64.0%88.9%96.3%3.7%
Expansion ratio ∥y∥/∥x∥\|y\|/\|x\| (median)allxx ≤50 tok51–200201–500>500
14.2×72×18×4.8×2.2×
Constraint typeall≤50 tok51–200201–500>500
Scene79.955.280.895.499.3
Temporal order56.87.752.495.699.4
Dialogue55.716.149.289.793.6
Style49.622.647.764.884.6
Camera43.97.739.963.897.3
Sound41.36.730.469.396.6
Lighting28.43.127.035.576.4
On-screen text11.70.95.417.744.6
Duration6.30.62.315.213.4
Aspect ratio4.70.92.911.64.1
# types per request (mean)3.81.23.45.67.1
zhen
Vocabulary size (types)78,36998,058
Hapax ratio32%24%
Requests with a unique opening 5-gram53.7%63.2%
Distinct-2 / 3 / 4, token-budget-matched (reconstructed xx)0.142 / 0.410 / 0.6470.312 / 0.578 / 0.801
Distinct-2 / 3 / 4, token-budget-matched (human-written prompts)0.161 / 0.445 / 0.6460.349 / 0.632 / 0.767

The table above shows that the reconstructed requests are diverse in vocabulary and phrasing, with diversity broadly comparable to that of human-written prompts.

SC-GRPO data. The SC-GRPO prompts are written by human annotators. We report their statistics below.

  • Token length: mean 428, median 223, max 5,597. Cumulative distribution ≤10: 0.5%, ≤50: 6.8%, ≤200: 46.3%, ≤500: 75.8%, ≤1,000: 89.5%, >1,000: 10.5%.
  • Language: 36 languages. zh 48.1%, en 43.0%, ru 2.0%, ar 2.0%, fa 1.1%, fr 0.7%, hi 0.4%, bn 0.4%, es 0.4%, pt 0.3%, ko 0.2%, ja 0.2%, ur 0.2%, de 0.1%, te 0.1%, id 0.1%, uk 0.1%, vi 0.1%, it 0.1%, tr 0.05%, mk 0.03%, tn 0.02%, ms 0.02%, bg 0.02%, th 0.02%, pl 0.02%, sw 0.02%, mr 0.02%, az 0.01%, hu 0.01%, tl 0.01%, nl 0.01%, sn 0.01%, da 0.01%, yo 0.01%, kk 0.01%.
  • Speech: 15.9% of prompts contain speech specified verbatim by the annotator, covering 25 languages: zh 53.7%, en 33.1%, ar 1.8%, ja 1.7%, nl 1.3%, fr 1.0%, es 1.0%, tl 0.8%, id 0.8%, de 0.7%, tr 0.5%, ms 0.5%, pt 0.4%, da 0.4%, cs 0.4%, fi 0.4%, pl 0.3%, ru 0.3%, it 0.3%, ko 0.3%, sv 0.2%, vi 0.2%, ro 0.1%, hu 0.1%, th 0.05%.

Evaluation on WanPEval.

WanPEval requests are written by human annotators to cover practical use cases of varying complexity. No selection was made with respect to request format or structure. Their statistics are:

  • Token length: mean 171, median 99, min 29, max 1,061. Cumulative distribution ≤50: 10.4%, ≤200: 69.1%, ≤500: 95.6%, ≤1,000: 99.6%, >1,000: 0.4%.
  • 12.4% of requests contain speech specified verbatim, in en 48.4%, zh 38.7%, ja 3.2%, fr 3.2%, tl 3.2%, de 3.2%.
  • The requests are written in en 56.2%, zh 36.1%, fr 2.0%, de 1.6%, ja 1.6%, ko 1.6%, and tl 0.8%. During evaluation, each request is also provided to the experts in a translation into their preferred language.
Reviewer_qcA3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper reframes text-to-video prompt enhancement as cinematic planning rather than simple prompt expansion. WanPE learns from real-world videos through video-grounded reverse supervised fine-tuning: it first produces hierarchical cinematic captions from videos and then reconstructs compatible user requests, so that the enhancer output aligns with video-grounded caption distributions. It further applies Semantic-Consistency GRPO with a nine-dimensional reward to penalize omissions, alterations, incorrect bindings, and temporal inconsistencies. The authors build WanPEval, a human-annotated 5–30 second testbed, and evaluate 4B–397B model variants with expert blind pairwise video comparisons, ablations, and cross-generator transfer experiments.

Critical strengths
  • The video-grounded reverse SFT design is novel and well motivated; ablations support that reverse construction outperforms forward rewriting and forward-target SFT.

  • The empirical effort is substantial, including multiple model scales, expert pairwise evaluation, duration/granularity breakdowns, and transfer to other generators.

  • SC-GRPO shows consistent gains in semantic consistency and downstream video preference, and the paper provides useful case studies of long-range binding failures.

Critical weaknesses
  • Comparison fairness is not fully controlled: WanPE is mainly evaluated with Wan3.0 as the downstream generator, while several commercial baselines are end-to-end APIs with native enhancers and generators, so generator quality and enhancer quality are not fully disentangled. 
  • The resource requirements are extremely high, including 512 GPUs, up to 397B models, and 1.05M videos. 
  • The SC-GRPO reward depends on an LLM evaluator over nine semantic dimensions, but the reliability, bias, and reward-hacking risk of this evaluator are not analyzed in sufficient depth. 
  • WanPEval contains 249 requests and expert evaluation is strong, but coverage across languages, cultures, genres, and non-English requests remains limited.
Questions for authors
  • Can the authors provide a stricter same-generator comparison between WanPE and native prompt enhancers, so that generator effects are separated from enhancement effects? 

  • How much do captioner errors and GPT-based request reconstruction errors propagate into the final enhancer, and were these errors quantified? 

  • How well does the nine-dimensional SC-GRPO reward correlate with human judgments, and is there any evidence of reward hacking or over-optimization?

  • To strictly isolate the effect of WanPE from the downstream generator quality, can the authors provide a controlled experiment where WanPE's enhanced prompts are fed into the exact same open-source generators used by the baseline rewriters?

5 comments
AuthorsOpenReview ↗

Response to Reviewer qcA3 (Part 1/3)

We thank the reviewer qcA3 for the thoughtful and positive feedback! We will try our best to answer all your questions. Please let us know if you still have further concerns, or if you are not satisfied with the current responses, so that we can further update the response ASAP.


W1 & Q1: Comparison fairness is not fully controlled / same-generator comparison between WanPE and native prompt enhancers

A: We thank the reviewer qcA3 for raising this concern about comparison fairness.

We address this concern from two perspectives: (i) native PEs of modern commercial video generation systems, (ii) pervious general-purpose PE methods.

(i) Native PEs of modern commercial systems.

  • Compare with the PE deployed in the production version of Wan3.0. To further validate the effectiveness of our native PE training approach, we report a previously completed internal evaluation. We invited 100 participants, including short-video creators, short-drama professionals, and general users, to test our system. We randomly sampled 2,400 prompts from their test inputs and compared WanPE with the native PE deployed in the production version of Wan3.0 (2026-08-27), following the human blind-evaluation protocol in Section 2.4. We compared two settings: (i) enhancing user prompts with a locally deployed WanPE-397B and submitting them to the Wan3.0 official API with the production PE disabled; and (ii) submitting the original user prompts to the same API with the production PE enabled. WanPE performed comparably to the PE deployed in production, with preference scores SS of 51.29% and 50.02%, respectively. This comparison demonstrates the effectiveness of our approach for training native prompt enhancers.

  • Compare with the native PEs of MiniMax-H3 and LTX-2.5. In Section 3.3, we compare WanPE against H3-Context-IR and LTX-2.5-PE on the same video generators, H3-Base and LTX-2.5-Base, respectively. The results in Table 2 demonstrate WanPE’s strong transferability.

  • For commercial systems such as Seedance 2.5/2.0, Kling 4.0/3.0, the native PEs are not independently accessible.

(ii) Pervious General-purpose PE methods

In the initial version, we did not compare against excellent prior PE methods such as RAPO, VPO, and Prompt-A-Video because we considered such a comparison unfair to them:

  • Different design targets. Prior PE methods were developed for previous-generation video models that natively supported only short-duration generation. Given their limited context capacity and relatively weak instruction-following capabilities of video models, prompt enhancement primarily served to enrich visual details. Their ability to generalize across diverse user inputs may also be limited.

Following the reviewer’s suggestion, we compare baselines on Wan3.0’s video generator with the 30-second subset of WanPEval, and on H3-Base with the 15-second subset due to its maximum generation duration, following the blind evaluation protocol in Section 2.4. Each baseline is compared with WanPE-9B (parameter is comparable to those of the trainable baseline PEs) in a separate 1v1 battle, and baselines are not compared against each other. The preference scores SS are as follows:

MethodActionAnim.SpeechAd.Sing & DanceDramaKnow.Overall
downstream generator: Wan3.0’s video generator
+RAPO00000000
+VPO000053.8501.64
+Prompt-A-Video07.140007.6902.46
+SCMAPR05.565.560153.8504.55
+WanPE-9B78.269093.5910077.5076.9292.8684.54
downstream generator: MiniMax-H3-Base
+RAPO01.727.14004.5501.72
+VPO8.333.4508.3312.509.092.505.17
+Prompt-A-Video06.9008.336.254.552.504.60
+SCMAPR8.3312.0714.298.3312.504.5559.20
+WanPE-9B79.1775.8673.2193.7582.8167.058077.59

The results show that WanPE-9B leads across all categories. We have also updated the demo page in the anonymous link, adding the videos and PE outputs of the baseline PEs on the original cases, which directly visualize the comparison with prior methods.

AuthorsOpenReview ↗

Response to Reviewer qcA3 (Part 2/3)

W2: The resource requirements are extremely high, including 512 GPUs, up to 397B models, and 1.05M videos.

A: We thank the reviewer qcA3 for raising this concern about resource. We clarify the three aspects below.

GPUs. The 512 GPUs were used to accelerate SFT and SC-GRPO training, not as a hard requirement. For a 4B model, 16 GPUs suffice for the full SFT and SC-GRPO pipeline.

Model size. The training method does not depend on model size. We apply the same recipe at 4B / 9B / 35B / 397B and observe consistent gains at every scale, which validates the generality of our method across model sizes. Table 3 show that the 35B model already outperforms all forward baselines on the semantic metrics of WanPEval.

Data. The corpus size is driven by coverage. For native PE training, generalization across diverse categories of user inputs is a primary consideration, including the languages, cultures, and genres raised in W4. Since reverse construction learns cinematic structure from real videos, this diversity must be present in the corpus itself. The corpus requires no manual annotation, as captioning and request reconstruction are fully automatic, so its cost is a controlable inference pass.


W3 && Q3: Reliability, bias, and reward-hacking risk of this evaluator are not analyzed in sufficient depth / correlate with human judgments, and is there any evidence of reward hacking or over-optimization?

A: We thank the reviewer qcA3 for raising this concern. We first address reliability and bias, then reward hacking and over-optimization.

Reliability and bias of the reward. Since the reward is obtained from LLM API calls and training is costly, we invested substantial effort in the reward dimensions and scoring rubric before SC-GRPO, finalizing them after 13 rounds of iteration. We then randomly sampled 200 cases, scored them with gpt-5.4, qwen3.7-max, and gemini-3.1-pro, and verified each score manually. Agreement with human judgment was 73%, 97%, and 98%, respectively. gpt-5.4 was overly strict on this task, while qwen3.7-max and gemini-3.1-pro performed comparably. For cost reasons, we chose qwen3.7-max as the training reward.

Reward hacking and over-optimization. Our evaluation protocol includes two checks.

  • The evaluation judge differs from the training reward. SC-GRPO is trained with qwen3.7-max, while the semantic consistency in Table 3(i) is scored by gemini-3.1-pro-preview, and by gpt-5.4 in the appendix. Consistent gains under both held-out judges indicate that the improvement is not overfitting to the training reward.

  • Human preference on downstream videos improves accordingly. In Table 3(ii), with the same Wan3.0 generator and only the enhanced conditions changed, 60 experts in blind pairwise evaluation raise the preference score from 42.70 to 49.69, with gains in all seven content categories. Over-optimization would typically raise the text-level reward without improving human preference.


W4: coverage across languages, cultures, genres, and non-English requests remains limited for WanPEval.

A: We thank the reviewer qcA3 for raising this concern.

We believe that generalization across diverse categories of user inputs is a primary consideration when developing native PEs for modern commercial video generation systems. Therefore, we include stylistically diverse data in WanPEval, as detailed below.

Genres. The 249 requests are balanced across seven content categories: Drama 44, Action 43, Speech 41, Animation 39, Sing & Dance 30, Knowledge 27, and Advertising 25. They cover genres including music performance 35%, slice of life 18%, sports 15%, sci-fi/cyberpunk 13%, education/science 12%, commercial 12%, historical/period 10%, fantasy/wuxia 9%, documentary/nature 9%, food/lifestyle 8%, horror/thriller 6%, romance 6%, crime/war 5%, and comedy 4%. Genre annotation is multi-label, so a single request may belong to multiple genres.

Cultures. 72.7% of requests are culturally neutral (e.g., cyberpunk, gym, product ads). Cultural markers are multi-label, as one request may contain elements from multiple regions. Measured over all requests, Chinese elements appear in 13.7%, Japanese in 3.6%, North American in 3.2%, European in 2.8%, Korean in 1.6%, and other regional elements in 2.8%.

Language of speech. 12.4% of requests contain speech specified verbatim, in en 48.4%, zh 38.7%, ja 3.2%, fr 3.2%, tl 3.2%, de 3.2%.

Language of request. The 249 requests are written in en 56.2%, zh 36.1%, fr 2.0%, de 1.6%, ja 1.6%, ko 1.6%, and tl 0.8%. During evaluation, each request is also provided to the experts in a translation into their preferred language.

AuthorsOpenReview ↗

Response to Reviewer qcA3 (Part 3/3)

Q2: How much do captioner errors and GPT-based request reconstruction errors propagate into the final enhancer, and were these errors quantified?

A: We thank the reviewer qcA3 for raising this question.

Captioner errors. Each caption is checked for structural completeness, timestamp validity, and consistency with the source video (Appendix B.3). 6.3% of captions fail and are regenerated.

Reconstruction errors. We use an LLM-as-judge to inspect every reverse-constructed pair (x,y)(x, y). A pair is treated as unqualified if xx introduces requirements unsupported by yy, or if xx incorrectly summarizes the content of yy. Such cases are harmful because they would teach the enhancer to invent unsupported constraints or to drop user requirements. After filtering, 89.1% of the pairs are retained. The semantic metrics of WanPE-397B-SFT before and after filtering are as follows:

MethodOverall↑5 seconds↑10 seconds↑15 seconds↑30 seconds↑Perfect↑Failure↓
Unfiltered73.170.268.578.970.726.942.6
Filtered75.573.270.081.573.429.736.9

Filtering improves overall semantic consistency by 2.4 points, raises Perfect, and reduces Failure, indicating that removing unqualified reverse-constructed pairs limits error propagation into the final enhancer.


Q4: experiment on exact same open-source generators used by the baseline rewriters.

A: We thank the reviewer qcA3 for this concern!

We have compared WanPE with the native PE (h3-context-ir and LtX-2.5-PE) of two open-source generators, H3-Base and LTX-2.5-Base, on the same generator(section 3.3). This is a particularly challenging comparison, as a native PE is naturally better aligned with its downstream video generator. WanPE leads in both comparisons.

At the time of the ICLR 2027 submission, MiniMax-H3 and the LTX series were the models we could identify that jointly offered native audio, native generation beyond 10 seconds, and independent access to both the native PE and DiT. These capabilities are important for studying modern cinematic PE under our evaluation protocol.

Following the reviewer’s suggestion, We have further added experiments with multiple prior PE methods on H3-Base, see our response to W1 & Q1.

Reviewer_DjxVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on prompt enhancement for text-to-video generation and proposes WanPE, aiming to address two key challenges in scaling modern video diffusion models: learning coherent cinematic planning across interdependent production components and preserving user-specified requirements throughout the unfolding plan. To this end, the authors introduce a video-grounded reverse construction strategy for building the supervised fine-tuning dataset and propose Semantic-Consistency GRPO to further improve semantic fidelity across shots and over time. In addition, the authors introduce WanPEval, a human-annotated benchmark for evaluating different WanPE variants and their downstream video generation performance.

Critical strengths
  • The paper explores the use of prompt enhancement to provide cinematic planning and structured textual guidance for text-to-video generation models.
  • The paper proposes a reverse construction strategy for building the SFT training framework. The authors provide empirical evidence demonstrating its advantages over conventional forward-based prompt enhancement, offering an interesting alternative for learning video-grounded cinematic planning.
Critical weaknesses
  • Insufficient systematic qualitative comparisons. Although the manuscript and its project page provide several prompt rewriting examples and generated video demonstrations, systematic qualitative comparisons of the downstream videos produced by different prompt enhancement methods are still lacking. In particular, it would be beneficial to present side-by-side comparisons of videos generated from the same user requests using different enhancement methods. Such comparisons would allow readers to better understand how the proposed rewriting framework affects the actual generation results.
  • Insufficient experimental validation of the prompt enhancement framework. Since the prompt enhancement model constitutes the central contribution of WanPE, a more comprehensive evaluation against existing prompt enhancement approaches, together with stronger evidence of its generalizability across different video generators, would further substantiate the effectiveness of the proposed method.
    1. Selection of baselines. Although the authors compare WanPE with Gemini 3.1 Pro-based forward rewriting in Table 2 (Exp. 2) and provide additional case studies involving other methods in Appendix J, a systematic large-scale quantitative comparison with existing prompt enhancement approaches is still missing. In particular, the current commercial model comparisons in Table 1 primarily evaluate end-to-end generation systems, making it difficult to isolate the contribution of the prompt enhancement component itself. A more controlled comparison under the same downstream video generator and evaluation protocol would be valuable. For instance, including task-specific approaches such as RAPO, VPO, and Prompt-A-Video, as well as LLM-based rewriting approaches using models such as GPT, Gemini, and DeepSeek, would provide a more comprehensive assessment of WanPE's effectiveness. Furthermore, providing the actual rewritten prompts produced by Gemini 3.1 Pro in the forward rewriting experiment would improve the transparency and reproducibility of the comparison.
    2. Generalization across different video generators. Although Table 2 (Exp. 3) demonstrates the transferability of WanPE to MiniMax-H3 and LTX-2.5 through format adaptation, the current evaluation remains relatively limited in terms of the number of downstream generators and the qualitative evidence provided. Providing more direct visual comparisons of the generated videos, together with evaluations on additional video generators, would help clarify the extent to which the learned cinematic planning capabilities generalize across different generation architectures and conditioning mechanisms, rather than serving merely as an extension module for Wan3.0.
Questions for authors
  • The necessity of a standalone prompt enhancement model. The paper argues that textual prompts play an increasingly important role in directing modern video generation models. However, state-of-the-art systems such as Seedance 2.5 increasingly support comprehensive multimodal reference conditioning and agentic planning mechanisms for fine-grained control and global content organization. Similarly, MiniMax-H3 provides its own prompt rewriting capabilities and various reference inputs to improve generation quality. Given these developments, it would be valuable for the authors to further clarify the intended role and practical necessity of a standalone, generator-agnostic prompt enhancement model, compared with relying on the increasingly sophisticated native conditioning and planning mechanisms of individual video generation systems. In particular, a clearer discussion of the advantages of WanPE over such model-specific approaches would strengthen the motivation and practical significance of this work.
  • Clarification of the video-grounded caption distribution. In Section 2.1, the authors formulate the objective of prompt enhancement as generating conditions that align with the video-grounded caption distribution. Accordingly, Section 2.2 constructs the training data by first captioning real-world videos to obtain yy, and then reconstructing a semantically compatible user request xx. However, to fully exploit the capabilities of a particular video generation model, the enhanced prompts should ideally align with the conditioning distribution of that model's own training data. These distributions may differ substantially across video generators, depending on their data sources, captioning models, and annotation or preprocessing pipelines. Consequently, the video-grounded caption distribution induced by the captioning model employed in WanPE may not necessarily match the actual conditioning distribution of the downstream generators. It would be helpful if the authors could further clarify the assumptions underlying this distributional alignment and discuss how the potential discrepancy between the constructed video-grounded caption distribution and the actual training caption distributions of different video generators might affect the effectiveness and generalizability of WanPE.
  • Coverage of user input distributions in reverse reconstruction. In Section 2.2, the authors construct training pairs by reconstructing user requests from video-grounded cinematic captions. While this strategy provides a principled way to obtain semantically compatible training pairs, it remains unclear how well the reconstructed requests cover the diversity of real-world user inputs. For example, users may provide highly detailed instructions that already specify shot composition, camera movements, dialogue, and temporal ordering, rather than concise high-level requests. Such inputs may differ substantially from the reconstructed requests used during training. It would be valuable if the authors could further discuss or evaluate the robustness of WanPE to these detailed user inputs, particularly when their original formats or structures differ from the model's preferred output format. This would provide additional evidence regarding the coverage of the proposed reverse construction strategy and the model's generalization to diverse user input distributions.
  • Further analysis of the experimental results and potential failure cases. There are two aspects that would benefit from further clarification:
    1. In Case 3 presented on the project page, the transition between Shot 6 and Shot 7 appears somewhat abrupt: the background changes from a train station to a rooftop, without an explicitly described transition connecting the two locations. The authors could clarify whether this is considered a failure case in terms of cinematic coherence and, if so, discuss the potential causes.
    2. In Table 2 (Exp. 3), WanPE-397B exhibits relatively weak performance on the Action category when paired with MiniMax-H3-Base, and on the Advertising category when paired with LTX-2.5-Base, compared with the corresponding native prompt enhancers. Further analysis of these category-specific results, ideally supported by representative video examples, would help clarify the limitations of the proposed method and the factors that influence its transferability across different generators.
6 comments
AuthorsOpenReview ↗

Response to Reviewer DjxV (Part 1/4)

We thank the reviewer DjxV for the thoughtful and constructive feedback! We will try our best to answer all your questions. Please let us know if you still have further concerns, or if you are not satisfied with the current responses, so that we can further update the response ASAP.

To better address the concerns, we first respond to the Questions For Authors, and then to the Critical Weaknesses.


Q1: The necessity of a standalone prompt enhancement model.

A: We thank the reviewer DjxV for this concern about necessity.

We would like to clarify that our primary research question is not how to build an external standalone prompt enhancer, but how native prompt enhancers could be trained for modern commercial video generation systems.

To the best of our knowledge, we are the first to openly investigate how to train native prompt enhancers in modern commercial video generation systems. These systems such as Seedance 2.5/2.0 and Kling 4.0/3.0 incorporate a native PE and a video generator. When their official APIs are called, user requests, including outputs from agentic planning mechanisms for creation that serve as pre-PE inputs, first pass through the native PE and then to the video generator [1]. However, the training methods and details of such native PEs remain largely undisclosed.

To further validate the effectiveness of our native PE training approach, we report a previously completed internal evaluation. We invited 100 participants, including short-video creators, short-drama professionals, and general users, to test our system. We randomly sampled 2,400 prompts from their test inputs and compared WanPE with the native PE deployed in the production version of Wan3.0 (2026-08-27), following the human blind-evaluation protocol in Section 2.4. We compared two settings: (i) enhancing user prompts with a locally deployed WanPE-397B and submitting them to the Wan3.0 official API with the production PE disabled; and (ii) submitting the original user prompts to the same API with the production PE enabled. WanPE performed comparably to the PE deployed in production, with preference scores SS of 51.29% and 50.02%, respectively. This comparison demonstrates the effectiveness of our approach for training native prompt enhancers.

In addition, we have compared against the native PEs of H3-Base and LTX-2.5-Base on their own native video generators in the initial version (Section 3.3), and the results demonstrate the generalizability of our native PE training method.

Regarding multimodal reference conditioning, PEs for video models can be categorized by their target generation tasks into T2V PE, I2V PE, and R2V PE [2]. Discussing how to support R2V tasks for native PE in modern commercial systems is the next steps in our work.

[1] Team, Kling, et al. "Behind Video Generation: Observations and Insights on Multimodal Understanding from Kling." arXiv preprint arXiv:2611.04916 (2026).

[2] Shao, Yawen, et al. "Towards Unified Evaluation of Prompt Enhancers for Video Generation." arXiv preprint arXiv:2610.11736 (2026).


Q2: Clarification of the video-grounded caption distribution.

A: We thank the reviewer DjxV for raising this concern about distribution.

In our experiments, Wan3.0’s multimodal captioner is used under authorization to construct training data for our native PE research. This supports alignment with Wan3.0 video generator’s training caption distribution.

We believe that a good native prompt enhancer needs to align with the distribution of the downstream video generator (section 2.1). Our distributional assumption is motivated by the following observation: for text-to-video generator, training pairs (x,v)(x, v) typically consist of real-world videos vv and corresponding textual descriptions xx produced by multimodal captioners [1][2]. If this downstream distribution is unknown (eg. the video model’s captioner is inaccessible), the caption distribution constructed by captioning real-world videos can serve as a reasonable proxy. Comparisons with the native PEs of H3-Base and LTX-2.5-Base on the native video generators (Section 3.3) also indicate that training on our real-world-video-grounded caption distribution supports transferability across video generation systems.

When applying our approach to train the native PE of a commercial video generation system, the multimodal captioner used in our data construction pipeline can be replaced with the captioner used to construct that generator’s training captions.

[1] Seedance, Team, et al. "Seedance 2.0: Advancing video generation for world complexity." arXiv preprint arXiv:2604.14148 (2026).

[2] Team, Kling, et al. "Behind Video Generation: Observations and Insights on Multimodal Understanding from Kling." arXiv preprint arXiv:2611.04916 (2026).

AuthorsOpenReview ↗

Response to Reviewer DjxV (Part 2/4)

Q3: Coverage of user input distributions in reverse reconstruction.

A: We thank the reviewer DjxV for raising this concern.

We have updated addition cases for you to review in the demo section 3.

We believe that generalization across diverse categories of user inputs is a primary consideration when developing native PEs for modern commercial video generation systems. Therefore, we include stylistically diverse data in both data construction and evaluation, as detailed below.

training data construction.

SFT data. Video-grounded targets yy are derived from real-world videos spanning ten content categories, each captioned with a category-specific instruction (Appendix B). The few-shot demonstrations used for reconstruction are drawn from a 2K human-written request pool stratified by category, covering granularities from one-sentence intents to detailed shot-level instructions. We report statistics of yy and xx over the full SFT set below.

  • Statistics of video-grounded targets yy:

    • Category and language: ten content categories; captions are written in Chinese (50.8%) or English (49.2%), and each xx is reconstructed in the language of its caption.
    • Length: min 275 tokens, mean 1,683, median 1,515, max 6,809.
    • Duration: median 14.0 s; ≤5 s 3.4%, 5–10 s 12.5%, 10–15 s 37.5%, 15–20 s 21.6%, 20–30 s 25.0%.
    • Speech: 60.3% of targets contain speech transcribed verbatim from the source audio, covering 25 languages: en 40.6%, zh 35.7%, ja 4.5%, ru 2.9%, es 2.1%, tr 1.8%, ko 1.7%, fr 1.5%, de 1.4%, tl 1.0%, it 1.0%, ar 0.8%, pt 0.7%, id 0.5%, nl 0.5%, vi 0.5%, th 0.5%, pl 0.4%, ms 0.4%, fi 0.4%, cs 0.3%, sv 0.3%, da 0.3%, hu 0.2%, ro 0.2%.
  • Statistics of reconstructed requests xx:

    • The four tables below show that xx spans a broad range of lengths and constraint types, with diversity comparable to human-written prompts.
Token lengthmeanmedianmax≤10≤50≤200≤500≤1,000>1,000
xx223931,8172.5%25.8%64.0%88.9%96.3%3.7%
Expansion ratio ∥y∥/∥x∥\|y\|/\|x\| (median)allxx ≤50 tok51–200201–500>500
14.2×72×18×4.8×2.2×
Constraint typeall≤50 tok51–200201–500>500
Scene79.955.280.895.499.3
Temporal order56.87.752.495.699.4
Dialogue55.716.149.289.793.6
Style49.622.647.764.884.6
Camera43.97.739.963.897.3
Sound41.36.730.469.396.6
Lighting28.43.127.035.576.4
On-screen text11.70.95.417.744.6
Duration6.30.62.315.213.4
Aspect ratio4.70.92.911.64.1
# types per request (mean)3.81.23.45.67.1
zhen
Vocabulary size (types)78,36998,058
Hapax ratio32%24%
Requests with a unique opening 5-gram53.7%63.2%
Distinct-2 / 3 / 4, token-budget-matched (reconstructed xx)0.142 / 0.410 / 0.6470.312 / 0.578 / 0.801
Distinct-2 / 3 / 4, token-budget-matched (human-written prompts)0.161 / 0.445 / 0.6460.349 / 0.632 / 0.767

The table above shows that the reconstructed requests are diverse in vocabulary and phrasing, with diversity broadly comparable to that of human-written prompts.

SC-GRPO data. The SC-GRPO prompts are written by human annotators. We report their statistics below.

  • Token length: mean 428, median 223, max 5,597. Cumulative distribution ≤10: 0.5%, ≤50: 6.8%, ≤200: 46.3%, ≤500: 75.8%, ≤1,000: 89.5%, >1,000: 10.5%.
  • Language: 36 languages. zh 48.1%, en 43.0%, ru 2.0%, ar 2.0%, fa 1.1%, fr 0.7%, hi 0.4%, bn 0.4%, es 0.4%, pt 0.3%, ko 0.2%, ja 0.2%, ur 0.2%, de 0.1%, te 0.1%, id 0.1%, uk 0.1%, vi 0.1%, it 0.1%, tr 0.05%, mk 0.03%, tn 0.02%, ms 0.02%, bg 0.02%, th 0.02%, pl 0.02%, sw 0.02%, mr 0.02%, az 0.01%, hu 0.01%, tl 0.01%, nl 0.01%, sn 0.01%, da 0.01%, yo 0.01%, kk 0.01%.
  • Speech: 15.9% of prompts contain speech specified verbatim by the annotator, covering 25 languages: zh 53.7%, en 33.1%, ar 1.8%, ja 1.7%, nl 1.3%, fr 1.0%, es 1.0%, tl 0.8%, id 0.8%, de 0.7%, tr 0.5%, ms 0.5%, pt 0.4%, da 0.4%, cs 0.4%, fi 0.4%, pl 0.3%, ru 0.3%, it 0.3%, ko 0.3%, sv 0.2%, vi 0.2%, ro 0.1%, hu 0.1%, th 0.05%.

Evaluation on WanPEval.

WanPEval requests are written by human annotators to cover practical use cases of varying complexity. No selection was made with respect to request format or structure. Their statistics are:

  • Token length: mean 171, median 99, min 29, max 1,061. Cumulative distribution ≤50: 10.4%, ≤200: 69.1%, ≤500: 95.6%, ≤1,000: 99.6%, >1,000: 0.4%.
  • 12.4% of requests contain speech specified verbatim, in en 48.4%, zh 38.7%, ja 3.2%, fr 3.2%, tl 3.2%, de 3.2%.
  • The requests are written in en 56.2%, zh 36.1%, fr 2.0%, de 1.6%, ja 1.6%, ko 1.6%, and tl 0.8%. During evaluation, each request is also provided to the experts in a translation into their preferred language.
AuthorsOpenReview ↗

Response to Reviewer DjxV (Part 3/4)

Q4: Further analysis of the experimental results and potential failure cases.

A: We thank the reviewer DjxV for this concern about detailed analysis!

We first address the reviewer’s concern regarding the transition between Shots 6 and 7 in Demo Case 3.

We agree that this transition is abrupt and regard it as a failure of end-to-end cinematic coherence. The PE preserves the user’s requested move from the train station to the rooftop, but does not provide sufficiently explicit guidance for connecting the two locations. The DiT also fails to establish visual continuity during rendering.

The contribution of the DiT can be isolated by holding the underlying WanPE plan fixed. In the Miyazaki-style Demo 5, Wan3.0 (Demo Section 1) produces a relatively natural transition between Shots 3 and 4 by maintaining character orientation and running motion, whereas H3-Base (Demo Section 2) produces a noticeably more abrupt cut from the same underlying plan. This comparison shows that transition quality depends on both PE planning and DiT rendering.

As DiTs become more capable of following detailed instructions, explicit transition guidance from the PE will become increasingly consequential. We will investigate this direction in future work.

We next analyze the lower scores of WanPE on the action and advertising categories of H3-Base and LTX-2.5.

To better understand these results, we break down the preference score SS by video duration.

Action (WanPE vs. H3-Context-IR, both on H3-Base):

5s10s15sOverall
WanPE22.22%22.73%50.00%29.63%
H3-Context-IR77.78%31.82%21.43%44.44%

Advertising (WanPE vs. LTX-2.5-PE, both on LTX-2.5-base):

5s10s15sOverall
WanPE10.00%25.00%42.86%27.50%
LTX-2.5-PE30.00%37.50%28.57%32.50%

In both settings, WanPE underperforms the native PE at short durations, which is expected given that a native PE is better aligned with its downstream video generator. As duration increases, WanPE’s cinematic planning advantage becomes clearer. This is consistent with our claim that prompt enhancement becomes increasingly important as the generation horizon grows (Line 358).


W1: Insufficient systematic qualitative comparisons.

A: Thank you for the concern on qualitative comparisons.

We have updated the anonymous demo page to include side-by-side comparisons with the additional cases you mentioned. In video generation quality, WanPE is substantially better than the other baselines.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

We are deeply grateful for your dedication throughout the review and rebuttal phases, as well as for the constructive feedback that has significantly enhanced the clarity and refinement of our work.

In response to the feedback, we have:

  • revised the manuscript and highlighted all changes from the initial draft in blue for your convenience.
  • updated the anonymous demo page with additional examples.
  • provided point-by-point responses to all comments.

We would like to summarize the key contributions of this work and highlight the experimental results incorporated during the rebuttal phase.


Overview

  1. To the best of our knowledge, we are the first to openly investigate how to train native prompt enhancers in modern commercial video generation systems. With the rapid advancement of DiTs, we believe that the PE should evolve from an optional auxiliary module into an indispensable director for fine-grained, holistic planning, warranting substantial research investment.

  2. Industry-scale empirical study and rigorous evaluation. Using industry-scale compute, we conduct a systematic scaling study across four model sizes (4B, 9B, 35B, and 397B), complemented by approximately 11K blind pairwise assessments by trained experts. These results establish the scalability of our training method, with gains across model scales consistently translating into improved downstream video quality.

  3. Real-world production impact. We have already licensed WanPE to the Wan 3.0 team. WanPE has been validated through large-scale deployment and real-world use and has received positive feedback from a substantial number of real-world users.


Experiments

  1. Comparisons with other PEs

    • Native PEs of commercial systems. In controlled comparisons where videos are generated using H3-Base and LTX-2.5-Base, WanPE outperforms the native PEs of MiniMax-H3 and LTX-2.5, respectively. Native PEs from other commercial systems such as Seedance and Kling cannot be accessed independently of their video generators. (Section 3.3, initial version)
    • Prior general-purpose PEs. We add comparisons with four prior general-purpose PE methods on two modern video generators, Wan3.0 and H3-Base. WanPE leads by substantial margins. (revised in Section 3.5)
    • Comparison with a strong forward baseline and additional ablations. We construct a strong forward-rewriting baseline, conduct a model-selection ablation, and apply the same SC-GRPO procedure to the forward-target SFT model. WanPE remains preferred under matched SFT and RL training, demonstrating the advantage of reverse construction. (Section 3.2, initial version, and revised in Appendices J.3 and K)
    • Directorial planning quality. Evaluation of enhanced prompts from 10 PEs across all nine Directorial Creation criteria shows that WanPE-397B achieves the strongest overall directorial-planning performance. (revised in Appendix H.3)
  2. Data quality and scaling. Reconstruction-based filtering improves semantic metrics, while training WanPE-397B-SFT on 25%, 50%, and 100% of the video-grounded corpus shows consistent improvements in semantic consistency as the data scale increases. (revised in Section 3.5 and Appendix D.1)


In-depth Analyses

  1. Evidence for generalization across diverse user inputs. Our training data and WanPEval cover diverse prompt lengths, intent granularities, languages, content categories, constraints, genres, and cultural contexts. (revised in Appendices D and F)

  2. Reward–human alignment. We report the agreement of different reward models with human judgments and select Qwen3.7-Max as the SC-GRPO reward for its high agreement with human judgments and lower cost. (revised in Appendix C.2)

  3. Caption-distribution alignment. We clarify that Wan3.0's multimodal captioner is used under authorization during the real-world-video-to-caption stage to construct the caption yy. This supports alignment with Wan3.0 video generator’s training caption distribution. (revised in Line 180 and Figure 2)

  4. Failure-case analysis. We analyze cases where WanPE underperforms H3-Context-IR and LTX-2.5-PE on their native generators. The duration breakdown highlights the growing importance of fine-grained PE control for long-video generation. (revised in Appendix I.2)

  5. Additional details. We clarify the annotation assignment and quality-control procedure used during evaluation, report the Gemini-3.1-Pro API parameters used for Forward Rewriting, and explain that 512 GPUs are used to accelerate training. (revised in Appendices G.3, K, and E)


Finally, we sincerely thank you for taking the time to provide constructive feedback on our work. Please let us know if you still have further concerns, we are happy to address them.

Best,

The Submission 8923 Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.