acceptodds
Under review as a conference paper at ICLR 2027

MoviEdit: Long-form Multi-Shot Video Editing Made Easy

Abstract

Long-form multi-shot video editing requires an intended edit to remain coherent across abrupt changes in scene, viewpoint, and subject appearance, while processing substantially longer visual context than conventional single-shot editors. We introduce MoviEdit, an end-to-end framework for scalable long-form multi-shot video editing that jointly addresses the challenges of data, modeling, and evaluation. To provide cross-shot supervision data at scale, we develop automatic data-generation pipelines spanning synchronized multi-view footage, real-world multi-shot videos, and generated multi-shot content, yielding 87K paired editing videos without manual annotation. Architecturally, MoviEdit employs a cascaded bidirectional-autoregressive design: a lightweight bidirectional stage first establishes a globally coherent editing draft over a spatiotemporally compressed representation, which is then autoregressively propagated and refined over the full video. We further introduce MoviEdit-Bench, a dedicated benchmark for systematic evaluation of long-form multi-shot video editing. Extensive experiments show that MoviEdit substantially outperforms existing open-source models and achieves performance comparable to strong proprietary models across text-guided and image-guided editing. Video results are available at this anonymous page: https://multi-shot-edit-anonymous.github.io/

Primary area:
generative models
Keywords:
video editing, multi-shot video, video diffusion models

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_fkEvOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MoviEdit combines a global, low-resolution bidirectional editing draft with autoregressive rendering to maintain edits across multiple video shots. It also introduces approximately 87K automatically generated training pairs and MoviEdit-Bench, reporting strong text-guided and edited-frame-guided editing results.

Critical strengths
  1. The shared-weight draft-and-render architecture directly addresses appearance drift and cross-shot consistency.
  2. Three complementary data-generation pipelines provide substantial supervision without manual pair annotation.
  3. Comparisons cover both editing modes, proprietary systems, and external short-form benchmarks.
Critical weaknesses
  1. MoviEdit-Bench contains only 64 generated videos, limiting evidence for generalization to diverse real-world cinematic footage.
  2. Heavy reliance on VLM judgments needs independent human validation, particularly for small differences against proprietary systems.
  3. Proprietary single-shot training data and the initialization checkpoint limit reproducibility despite the planned pipeline release.
  4. Full-sequence bidirectional draft attention remains a scaling bottleneck, while demonstrated extrapolation to 25 seconds provides limited evidence for substantially longer videos.
Questions for authors
1 comment

Dear Reviewer fkEv,

Thank you again for your thoughtful review and constructive feedback.

As the discussion period is approaching its end, we would like to kindly follow up on our rebuttal. We have made efforts to address your concerns, particularly by adding a human preference study, quantitative evaluations of temporal generalization and inference scalability, and further clarifications on benchmark design and reproducibility.

We would greatly appreciate hearing whether our responses have addressed your main concerns. If any questions or concerns remain, we would be happy to provide further clarification.

Thank you again for your time and consideration.

Reviewer_dhAgOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper presents an end-to-end framework for scalable video editing tasks. The paper argues that existing video editing methods are typically designed for short (a few seconds) and single-shot videos and mostly fail when extended to long-form and multi-shot videos. The paper identifies three core challenges in existing video editing systems: data, modeling, and evaluation. To address these challenges, the paper proposes three components, each targeting one of these challenges. First, a scalable data generation pipeline is developed that combines multi-view-aware image editing and single-shot video editing to construct paired training data with explicit subject correspondences across shots. Then, using this data, the authors train a cascaded video editing architecture in which a bidirectional editing stage first constructs a globally coherent edit draft followed by a causal autoregressive stage (similar to existing methods) that propagates the edit across the full input video sequence. Finally, the authors develop an evaluation suite specifically designed for long-form multi-shot video editing and compare their results with both proprietary and open-source video editing models. Extensive experiments on this newly developed benchmark, as well as existing benchmarks for short single-shot video editing, show that, across both text- and image-guided editing, the proposed method outperforms existing open-source video editing approaches and performs comparably to, and sometimes better than, proprietary video editing systems.

Critical strengths
  • The paper is well-written and easy to follow.

  • All three proposed components, especially the data generation pipeline and the evaluation protocol and data for video editing, can be valuable for further research in this field.

  • The results across extensive video editing tasks, compared against a wide range of baselines, including both proprietary and open-source methods, demonstrate strong progress in the field. The proposed method outperforms the open-source baselines and performs on par with the proprietary baselines.

  • The paper presents extensive ablations and experiments to demonstrate the effectiveness of the different proposed components.

  • The Draft Editor, a lightweight full-sequence diffusion module, is a simple but effective idea for overcoming the limited global context of chunk-wise autoregressive video generation models. It also helps remove side effects such as appearance drift and color saturation that occur in existing autoregressive video models as shown in Figure 3 of the paper.

Critical weaknesses
  • While the authors propose different optimization techniques in Sec. 2.3 to accelerate inference, and I appreciate their efforts in this regard, there is no comparison of inference time between the proposed method and the baselines, including both proprietary and open-source methods. Therefore, it is not clear how fast the reported inference time of 11.8 seconds in Table 4 is compared to existing methods. I would recommend reporting the inference time of the baselines as well, as this would give readers better insight into the speed of the proposed approach and help position it among existing video editing methods from an efficiency perspective. Also, regarding memory usage during inference, since the proposed method includes an additional Draft Editor module compared to existing chunk-wise causal autoregressive diffusion models, it likely requires more memory during inference. I would recommend reporting and comparing the inference memory usage of the proposed method and the baselines as well.

  • While the paper provides extensive qualitative and quantitative comparisons with baselines, it lacks a user preference study. Since the generated outputs are edited videos, it would be valuable to ask users to evaluate different aspects of the editing task, such as whether the requested edit is properly performed and whether the method preserves temporal and spatial consistency across video frames. Comparing the proposed method with existing methods across these aspects, with a sufficient number of participants, could further verify the effectiveness of the proposed approach for video editing tasks.

  • The paper claims that the proposed method can also work with videos that are longer or contain more shots than those used to train the MoviEdit model. Although this is a positive aspect, the paper demonstrates this capability only by testing the model on videos with more frames (longer than 15 seconds) and more shots, as shown in Figure 13. However, there is no supporting evidence or explanation (at least I could not find any justification, discussion, or specific component) that clarifies why the method can generalize to such videos and how it achieves this capability.

  • Regarding the Render Editor module in Sec. 2.2, my understanding is that the main (and possibly the only) difference between the proposed chunk-wise autoregressive generation approach in this work (Equation 6) and existing approaches (Equation 4) lies in the conditioning signals from the Draft Editor module that are fed into the Render Editor. If this understanding is correct, I would recommend explaining more clearly how these conditioning signals are architecturally and practically incorporated into the Render Editor, as the current presentation is not clear on this point. Moreover, the current presentation makes it seem as though chunk-wise autoregressive generation itself is proposed in this work, whereas the main change appears to be the conditioning signals applied to this type of model. I would recommend clearly distinguishing this difference and explicitly identifying the novel contributions relative to existing methods, as this would improve the clarity of the manuscript and better highlight the main contributions of the work.

  • Minor comment: In some tables of the paper (e.g., Tables 1, 2, and 5), the metric names are densely packed into the columns, particularly with the arrows indicating whether higher or lower values are better. I would recommend improving the formatting and spacing to make these tables easier to read.

Questions for authors
  • Regarding MoviEdit's ability to perform editing beyond the training horizon, which I also pointed out in the weaknesses, could the authors explain why this is possible and clarify whether any specific design choices, components, or contributions have been developed to enable this capability?
1 comment

I appreciate the authors' efforts in addressing my comments. The authors have properly addressed all of my concerns. Therefore, I increase my score to 4 (clear acceptance).

Reviewer_QvsgOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MoviEdit, which edits videos with several shots using either a text instruction or an edited reference frame. The authors build 87K paired editing videos with three data pipelines. The model first makes a compressed draft using bidirectional attention. It then edits the full video chunk by chunk using the draft. Distillation reduces the draft and render stages to 12 and 4 steps. The new MoviEdit-Bench has 64 generated source videos and 128 editing cases. V2V overall success is 0.790, compared with 0.764 for Aleph 2.0. Full-benchmark IV2V results are 0.786 and 0.781. The reported denoising time is 11.8 seconds for 361 frames at 540p on 8 * H100 GPUs.

Critical strengths
  1. The data pipeline edits the first frames of all shots together, then carries each edit through its shot. This is a practical way to make training pairs with consistent edits across shots. The localization and semi-pixel-aligned data cover cases that standard editing pairs may miss. The benchmark checks edit location, subject preservation, and cross-shot consistency separately. This helps show why an edit fails.

  2. Tables 1-2 show large gains over the tested public models for both text-guided and reference-frame-guided editing. Results are also close to the best closed-source systems. The authors test FiVE-Bench and VPBench as well as their own benchmark. Table 4 reports the time spent on the draft and render stages and shows how windowing, caching, and distillation reduce it.

Critical weaknesses
  1. Figure 3 shows examples with draft conditioning, but does not compare quality scores for causal-only editing, the two-stage method, and full attention using the same backbone and training data. Table 4 measures running time, while the data ablations in Appendix D.2 also show only examples. So it is hard to tell how much each part helps. The draft still uses full-sequence bidirectional attention, and its cost grows with video length. Two 25-second examples are not enough to show how well the method handles longer videos.
Questions for authors
  1. How much does the draft improve quality over a causal-only model trained with the same backbone and data? Please report quality after 12/4-step distillation too. What are the model size and spatial draft compression? How do memory use and running time change beyond 15 seconds?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.