acceptodds
Withdrawn from ICLR 2027

Extending the Temporal Horizon of Video Editing via Self-edit Adaptation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent video editing models are trained on datasets of mostly short videos, which limits their temporal horizon, i.e., the temporal range over which their editing capability is reliably maintained. Although a long video can be processed at once, we find that edits succeeding on short inputs break down into visual artifacts or inaccurate modifications. Editing it in shorter segments instead avoids this breakdown but yields a different edit in each segment, which poses a trade-off between editing quality and temporal consistency. To resolve this, we propose a simple yet effective method, coined , which adapts the model to its own edit that stays within the horizon yet spans the full length of the video. Specifically, we compose consecutive frame groups sampled across the video into a short adaptation clip, covering diverse temporal states despite inter-group discontinuities. Since this discontinuity makes the clip harder to edit even within the horizon, we edit it with source-aligned generation, which keeps the editing trajectory aligned with the source clip. We then train a LoRA on the adaptation clip alone, with its edited latent as the flow-matching target. Finally, the adapted model edits the full video with the same source-aligned generation, avoiding a mismatch between the adaptation target and the inference trajectory. Consequently, without any additional training data, HESA extends the temporal horizon of pretrained video editing models, improving both editing quality and temporal consistency on long and dynamic videos.

Primary area:
generative models
Keywords:
Video Editing, Model Adaptation, Long-Video Editing

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_e4rYOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Horizon Extension via Self-edit Adaptation (HESA), a framework designed to extend the temporal horizon of pretrained video editing models. HESA constructs a short adaptation clip from consecutive frame groups sampled across the input video and trains LoRA parameters using the clip’s source-aligned edited latent as supervision, while keeping the pretrained backbone frozen. The reported results show improved source preservation and temporal alignment on long and dynamic videos without requiring additional training data.

Critical strengths

HESA addresses a practical limitation of pretrained video editing models: maintaining editing quality when processing videos substantially longer than those seen during training. Its self-edit adaptation strategy is straightforward and uses the model’s own edited short clip as supervision, without requiring additional training data.

Critical weaknesses
  • Limited novelty and insufficient differentiation from prior adaptation-based editing: Video-specific model adaptation has been explored in Tune-A-Video and Dreamix, while I2VEdit and LoRA-Edit further demonstrate LoRA-based adaptation for video editing. In particular, LoRA-Edit uses edited reference frames as supervision during fine-tuning, providing a relevant precedent for learning desired editing outcomes through model adaptation. Although HESA differs by using a self-generated adaptation clip to extend the temporal horizon, the methodological advance over these prior adaptation strategies is not sufficiently established. A more explicit comparison of the supervision targets, adaptation procedures, and how the learned edits transfer to other frames would help clarify the significance of the proposed contribution.

  • Insufficient experimental validation: The main evaluation is centered on extending VideoCoF, with an additional applicability study on LucyEdit, but does not establish how HESA compares with dedicated long-video editing approaches. Comparisons with methods such as AnchorSync and AdaFlow would be particularly informative, as they explicitly address long-range editing consistency through alternative mechanisms. In addition, the independent short-segment baseline illustrated in Fig. 1 should be evaluated quantitatively on the same test set, rather than only qualitatively. The evaluation is also limited to 24 videos and 72 video–prompt pairs. Broader validation on an existing long-video benchmark, such as LongV-EVAL, would strengthen the evidence. These comparisons should consider editing faithfulness, source preservation, temporal consistency, and end-to-end computational cost, including per-video adaptation.

Questions for authors
  • Temporal discontinuity in adaptation clip construction: The proposed grouping preserves local continuity within each group, but temporal jumps remain between groups, as acknowledged in Sec. 2.1. Could the authors clarify whether the VAE’s temporal receptive field spans these boundaries and how such discontinuities affect the encoded latents? While Table 4 compares downstream editing performance, a direct analysis of VAE reconstruction quality or boundary artifacts would help substantiate the proposed explanation for the benefit of consecutive grouping.

  • Generalization to different video lengths: Most experiments are conducted on 301-frame videos. Could the authors provide quantitative comparisons on videos of different lengths, such as approximately 100 frames and more than 500 frames, to demonstrate how the effectiveness of HESA varies with the temporal horizon?

  • Comparison with segment-wise short-clip editing: Could the authors provide a full-sequence quantitative comparison against the independent 33-frame segment-editing baseline illustrated in Fig. 1? The evaluation should assess the complete 301-frame videos reconstructed from the edited segments, including editing quality and consistency across segment boundaries. Please also clarify whether the 33-frame results in Table 2 correspond to this setting or to evaluation on short clips alone.

  • Computational overhead of per-video adaptation: HESA requires additional optimization for each input video, reportedly taking approximately 15 minutes on an NVIDIA B200 GPU. Could the authors provide an end-to-end runtime and memory comparison against existing approaches, including sequential short-clip editing, to better justify the practical benefits of the proposed adaptation strategy?

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

This paper studies instruction-based editing of long videos. Recent video editing models such as VideoCoF are trained on short clips (33 frames), so their temporal horizon is limited: when they edit a long video in one pass, an edit that works on a short clip can fall apart into artifacts or wrong changes. Editing the video in short segments keeps the quality, but each segment is edited independently and can get a different edit, for example a different object in each part of the video. The paper proposes HESA (Horizon Extension via Self-edit Adaptation), which gets one consistent edit across the whole video by teaching the model its own edit, without new training data. HESA builds a 17-frame adaptation clip from four groups of consecutive frames spread across the video, so the clip covers the video's different states while staying within the training length. The frozen editor edits this clip with FlowAlign-based source-aligned generation to keep the edit close to the source, a small LoRA (rank 4, 500 steps) is trained on the clip alone with its edited latent as the flow-matching target, and the adapted model then edits the full video in one pass with the same source-aligned generation, so the model never sees the full video during adaptation. The authors claim three contributions: horizon extension through adaptation to the model's own edit of a short clip, results on two editors (VideoCoF and LucyEdit) at more than 9× the training length, and an analysis of what makes adaptation work. On 72 video-prompt pairs of 301 frames, HESA raises the overall V2V-Bench score from 0.68 to 0.74 (+8.3%), above the 0.72 that VideoCoF reaches on 33-frame clips. The authors also highlight larger gains on dynamic ad videos (9.7% vs. 7.0%), the need for both broad coverage and locally consecutive frames in the adaptation clip, and the complementary roles of the two components: source alignment keeps the video's structure but weakens the edit, and adaptation restores edit strength, raising VQAScore from 0.56 for VideoCoF to 0.60.

Critical strengths
  1. HESA is a simple and practical way to extend a pretrained video editor to long videos. It needs no extra training data and no change to the base model: the model edits a short clip within its training range, and a small LoRA stores that edit. The model never sees the full video during adaptation, yet the qualitative results show one coherent edit across 301 frames, more than 9× the training length, including ad videos with shot cuts and large scene changes. The method is applied to two different editors (VideoCoF and LucyEdit), so it is not tied to one model.

  2. The design choices are well reasoned and checked. The adaptation clip uses groups of consecutive frames because the Wan VAE compresses consecutive frames together, and spreading the groups covers the different states of the video. Using the same source-aligned generation at adaptation and inference also matters: with standard generation at inference, the Semantic Alignment gain over VideoCoF drops from 77.5% to 32.2%. The same table shows that the adapted model alone, with standard generation, raises VQAScore from 0.51 to 0.60, which suggests that the LoRA itself helps the edit carry over to the full video.

  3. The analysis is transparent about what each part does and where the method fails. Table 5 separates source alignment from adaptation and reports that source alignment weakens the edit (VQAScore 0.56 → 0.47), while adaptation restores it (0.60). The authors note that the CLIP-based Transformation Quality metric misses this drop and add VQAScore to measure it. They also show that HESA depends on the base editor's edit of the adaptation clip, with a failure case (Fig. 11) and weaker results on LucyEdit (Fig. 12), and they note that all methods struggle under the large scene change in Fig. 7.

Critical weaknesses
  1. Most of the headline gain comes from source alignment, and the key baselines are missing. Tables 1–3 compare HESA, which uses source-aligned generation, against plain VideoCoF. In Table 5, VideoCoF with source alignment alone scores as high as or higher than HESA on Temporal, Structural and Semantic Alignment (0.73 / 0.83 / 0.79 vs. 0.71 / 0.81 / 0.78), so the LoRA adaptation mainly restores edit strength. This comparison uses only a 24-pair subset, and the only check of adaptation without source-aligned inference (Table 6) uses 12 pairs. The paper also gives no numbers for segment-wise editing, the trade-off that motivates it , and does not compare with long-video editing methods it cites, such as AnchorSync, or with training-free length extrapolation for video DiTs, such as RIFLEx (ICML 2025). The claim that HESA surpasses the 33-frame VideoCoF score (0.74 vs. 0.72) also compares a method with source alignment to one without. This matters because these comparisons decide whether the adaptation itself, the paper's main contribution, improves long-video editing.

  2. The metrics favor edits that change little, and the evaluation is small. Most V2V-Bench scores measure closeness to the source, so a weak edit can score well: LucyEdit has a VQAScore of only 0.32 and a Semantic Alignment of 0.78, while VideoCoF has 0.56 and 0.48 (Table 5). Transformation Quality stays at 0.60 for every method, and VQAScore is reported only on the 24- and 12-pair subsets, so the full-set results in Table 1 do not show whether HESA performs the edit better. V2V-Bench is a new arXiv benchmark without a reported link to human judgment, and the paper has no user study. The main evaluation uses 72 pairs from 24 videos, the ablations use 24 or 12 pairs, and no results include seeds or confidence intervals, while the clip-design differences in Table 4 are only 0.01–0.03. This matters because the paper's claim of better editing quality rests on these results.

  3. One more concern is cost. HESA trains a new LoRA for every video and prompt, which takes about 15 minutes on a B200, and source-aligned generation runs two trajectories at each step. The paper calls the procedure "computationally efficient" (Sec. 2.2), but it reports no inference time and no comparison with single-pass or segment-wise editing. For a practical editing tool, users will want to know this cost.

Questions for authors
  1. Could you report VideoCoF with source alignment and segment-wise editing (separate 33-frame segments) on the full 72-pair set?
  2. Could you compare with at least one other way to handle long videos, such as AnchorSync, using the edited adaptation clip as keyframes for propagation, or a training-free length-extrapolation method such as RIFLEx?
  3. Could you report VQAScore or a small user study on the full set, with seeds or confidence intervals for the main tables? This would show that HESA performs the edit, beyond keeping the source.
  4. What is the total time per video-prompt pair (adaptation plus inference), compared with single-pass and segment-wise editing? Can one LoRA be reused for several prompts on the same video?

Clarifications:

  1. Sec. 3.4 gives VideoCoF a Temporal Alignment of 0.62 on non-dynamic and 0.48 on dynamic videos, with 12 of the 24 videos counted as dynamic. The overall value in Tables 1 and 2 is also 0.62, while an average of about 0.55 would be expected. How is each number computed?
  2. How sensitive is HESA to the alignment strength γ, the number of LoRA steps and the number of frame groups? The 17-frame clip covers less than 6% of a 301-frame video, so how does HESA behave on much longer videos or videos with many shot cuts?
Reviewer_qKGROpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents HESA to address the long video editing method by model self-adaptation on a shorter clip and propagation of the edits to the whole length. Specifically, it first uses source alignment methods to edit the uniformly sampled short clip from the long source video, and then uses the edited clip to train the LoRA on the backbone. After this self-adaptation, it again uses this model to perform long video editing using the source alignment method for the original long video. It incorporates the editing knowledge into the backbone with the LoRA, together with source alignment, to achieve editing of the long video.

Critical strengths

It tackles a meaningful problem of long video editing, utilizing self-adaptation on the short clip to achieve editing for the whole length. This idea itself is reasonable.

Critical weaknesses

Novelty is very limited. Source alignment editing and LoRA-based editing are not new; many methods use these approaches to achieve image/video editing, such as FlowAlign and LoRA-Edit. This paper simply combines them and add incremental modifications; these contributions seem insufficient for acceptance to me.

The performance also seems limited, as shown in Table 5. The complete version (with both the source alignment and model adaptation) achieves lower scores on the first three metrics and only improves edit faithfulness. It shows that the performance gain mainly comes from the source alignment, and the proposed method reaches a trade-off point. This weakens the main contribution of the paper, making the central claim less convincing. Furthermore, It is uncertain whether the performance could be improved by changing the editing extent of source alignment alone.

Given these concerns, I lean toward a rejection.

Questions for authors

NA

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

Related papers

Discussion (0)

Sign in to comment.