acceptodds
Under review as a conference paper at ICLR 2027

OpenWan-Light: High-Quality Data Enhancing and Accelerating Video Generation

Abstract

Recent advances in video generation have greatly improved visual quality and temporal coherence, but the growing scale of modern models makes inference increasingly expensive, limiting their use in real-time applications. While models and acceleration algorithms are becoming increasingly open, the data used to train high-quality accelerated video models remain comparatively underexplored. We introduce a 450K real-video dataset centered on rich and diverse motion while preserving fine visual detail and strong text alignment, curating from 4.55M clean, standardized candidate clips. Our data-curation pipeline combines standard filtering and agentic VLM-based review to identify and balance dynamic video content. To demonstrate its effectiveness, we further build a complete acceleration pipeline on Wan2.2 A14B, combining supervised fine-tuning and few-step distillation with a generator-matched one-step refiner for high-resolution detail recovery. Without quantization or sparse attention, our system achieves acceleration over the reference generation pipeline and generates a 5-second 480p video in approximately 4.9 seconds on single NVIDIA H100, while maintaining strong generation quality and user preference. We will publicly release the dataset, model checkpoints, and complete training and inference code to support reproducible research on efficient and dynamic video generation.

Primary area:
datasets and benchmarks
Keywords:
Video datasets, Video data curation, Few-step distillation, Real-time video generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_c1rkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes OpenWan-Light, which consists of a high-quality, motion-oriented dataset and an open acceleration pipeline. A scalable two-stage data curation framework is developed, combining lightweight standard filter operations and a VLM-based review process. The authors accelerate the Wan2.2 model with the proposed acceleration pipeline on the curated dataset. The results deliver a better VBench total score against the original teacher model with a 76$\times$ speedup.

Critical strengths
  1. The paper presents a practical end-to-end framework for accelerating video generation. It combines a standardized motion-centric video dataset with a complete Wan2.2 acceleration pipeline. The results deliver a better VBench total score with a 76$\times$ speedup.
  2. The generator-matched refiner is well-motivated and practical. Figure 5 also shows significantly better performance with the generator-matched degradation.
Critical weaknesses
  1. The unique contribution of the curated dataset compared with existing datasets needs further clarification. The paper doesn't provide a fair comparison of fine-tuning or distillation results on different datasets.
  2. The effectiveness of motion-centered selection remains unclear. First, the improvement in mean optical flow between the curated dataset and the pre-curation dataset is relatively small (from 3.28 to 3.36). Second, the proposed method achieves the lowest dynamic degree metric in VBench among the baselines (77.22 for the Wan2.2 teacher model; 71.94 for OpenWan-Light; 88.06 for TurboDiffusion; 75.56 for LightX2V).
Questions for authors
  1. Could you provide controlled fine-tuning comparisons using matched numbers of clips and matched training budgets for existing datasets and the proposed curated set? Results beyond distillation would also be helpful. For example, could you provide comparison results for direct fine-tuning on the proposed dataset and other datasets?
  2. How effective is the motion-centered selection pipeline? The median of the per-clip mean RAFT optical flow increases only from 3.28 before curation to 3.36 after curation. Could you explain why the improvement is limited? It would also be useful to distinguish and report improvements in subject motion from those in camera motion and artifacts. For example, could you locate or segment subjects in the video and calculate their locomotion?
  3. Please clarify how each baseline was implemented. Were LightX2V and TurboDiffusion evaluated using their official checkpoints, or were they retrained/fine-tuned on the proposed 450K curated dataset? If the baselines were not trained on the same data, please provide a controlled comparison under matched training data and compute budgets.
  4. Why does the final model have the lowest Dynamic Degree among all compared methods? OpenWan-Light obtains 71.94, compared with 77.22 for the Wan2.2 teacher model, 75.56 for LightX2V, and 88.06 for TurboDiffusion. Could you explain whether distillation or refinement suppresses motion amplitude or diversity?
Reviewer_Cc1COpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work mainly proposes a text-to-video dataset, consisting of a candidate pool of 4.55 million clips and a 450K subset with higher motion intensity. The footage comes mainly from Pexels. The curation applies standard filters followed by a VLM review pass, which checks visual quality, camera motion, slow motion, and shot cuts, and also generates captions. To show that the data is useful, the authors build a training pipeline on top of Wan 2.2 T2V 14B that includes SFT, distillation for acceleration, and super-resolution refinement. It generates a 5-second 480p video in about 4.9 seconds on a single H100, with quality improved over the original teacher.

Critical strengths
  1. It contributes a valuable training dataset to the video generation field and validates its effectiveness in improving motion quality. Parts of the data-processing pipeline are a useful reference for others. Table 1 lays out the source composition and each filtering stage, making it easy for readers to understand what the dataset actually contains.

  2. Through the combination of SFT, distillation, and refiner-based generation, it achieves more efficient generation. In addition, the way the refiner's training inputs are constructed is worth noting: instead of using synthetic blur, degraded latents are passed through the deployed few-step model for one step (Section 4.3), so the refiner learns the failure modes the student model actually exhibits. The approach is simple and would probably transfer to other cascaded architectures.

Critical weaknesses
  1. The contribution of the data is not convincingly demonstrated. The paper never runs the same SFT-plus-distillation pipeline on a different set of 450K clips (whether 450K clips randomly sampled from the pool, a public dataset such as OpenVid, or Pexels clips that skip the VLM stage), so I cannot tell what this filtering actually brings. The only evidence is Figure 4, which compares two prompts against the official teacher, and that comparison conflates the effect of the data with the effect of low-resolution SFT itself, which the paper itself states is necessary anyway.

  2. The evaluation is relatively weak. The abstract and Section 1 claim the highest user preference among the compared systems, yet there is no user study anywhere in the paper. The VBench gains in Table 2 are marginal, only 0.11 above LightX2V and 0.24 above the teacher, which in my view amounts to a tie.

Questions for authors
  1. Could you validate the same pipeline on other data, for example by training on 450K clips randomly sampled from the pool, or on 450K OpenVid clips, and report a motion metric plus a small-scale human preference test? If the filtered data shows a clear advantage, I will raise my score. This is the experiment I most hope you will prioritize.
  2. Please clarify which VLM, and of what size, is used for data filtering. On a few hundred clips, what is the agreement rate between its pass/reject decisions and human annotations?
  3. Figure 1 states that the 4.9 s generation time includes T5, but the latency breakdown in Figure 3 already adds up to 4.9 s without a T5 entry. Could you clarify whether the 4.9 s includes T5? How is GPU memory managed for the three 14B DiTs? Does the teacher's 40 NFE include CFG?
  4. Could you explain why there are no user study results to support the experimental conclusions?
Reviewer_WxJQOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper builds a 450K video-text dataset focused on rich motion and uses it to fine-tune and distill Wan2.2. The system first generates a low-resolution video in four steps, then uses a one-step refiner to recover higher-resolution details. The refiner is trained with degraded inputs produced by the base model to better match the errors seen during inference.

Critical strengths
  • Curating real videos with rich motion and accurate captions is useful. The planned release of the data, models, and code could help other researchers.
  • The reported speed is promising: about 4.9 seconds to generate a five-second 480p video on one H100, with competitive VBench scores.
  • Training the refiner on degradations produced by the base model makes sense. The examples show improvements in local structure and detail.
Critical weaknesses
  • My main concern is the novelty of the acceleration pipeline. Generating at low resolution and then adding a refiner has already been explored in works such as FlashVideo and Imagen Video. FlashVideo also trains its second stage with simulated degradations to repair artifacts from the first-stage output. Although these works are cited, I would like a clearer explanation of what is new here, especially in the refiner training.

  • The connection to motion-rich data is also unclear. Such data may help the base model learn better actions, but this does not explain why the two-stage pipeline itself is a new contribution. The current examples do not separate the effect of motion selection from low-resolution adaptation and additional training.

  • Please compare the curated data with an equally sized, quality-matched dataset using the same training budget and pipeline. A comparison with existing refinement methods on the same low-resolution inputs would also help. These tests would make it easier to see whether the main contribution comes from the data, the refiner training, or their combination.

Questions for authors
  • What is new in the two-stage pipeline? Low-resolution generation followed by high-resolution refinement has already been explored in FlashVideo and Imagen Video. These works are cited, but the differences are not explained clearly. Which parts of the proposed pipeline go beyond these existing approaches?

  • What is the connection between motion-rich data and this pipeline? The data may help the model learn better actions, but the acceleration design seems general. Please compare the curated data with an equally sized, quality-matched dataset under the same training budget and pipeline. This would help separate the benefit of motion selection from low-resolution adaptation and additional training.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.