acceptodds
Under review as a conference paper at ICLR 2027

InfinityEdit: Infinite Video Editing with a Lightweight Edit-Ignition Adapter

Abstract

Existing methods have improved instruction-based video editing. However, most of them adopt an in-place editing pattern. They align the edited video with the given source clip over a fixed time span. This pattern fails for open-ended streams, e.g., restyling a live game or applying a camera move to an ongoing shot. In such cases, edits must extend to future frames as they arrive, rather than be applied to a static input clip. In this paper, we study this setting and name it infinite video editing: given a preceding segment and an edit request, a model must generate the next segment that continues the stream while applying the edit. This process repeats as an unbounded sequence of edit instructions arrives. This task brings two challenges: the edit must be a faithful continuation rather than a rewrite, and quality must remain stable as edits accumulate. To address them, we propose InfinityEdit, a lightweight edit adapter that equips a streaming video generator with unbounded editing ability. The adapter contains three attention modules. History cross-attention guides the denoising frames using the input frames. Temporal causal self-attention keeps temporal cues flowing only from earlier frames to later ones. Edit cross-attention injects the edit request into generation. During inference, the adapter is activated only in the chunk where an edit request arrives. Subsequent chunks are generated by the original model with a reset anchor frame. This scheme applies the edit while preserving the original model's infinite generation ability. Experiments show that InfinityEdit faithfully continues the stream under each edit, and stays stable over unbounded edit sequences.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
video editing, video generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_bNFUOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce the task of Infinite Video Editing, which aims to generate future video segments according to editing instructions while maintaining temporal continuity with the preceding content. To address this task, the authors propose InfinityEdit, a lightweight editing adapter built on a frozen Helios backbone. During inference, the adapter is activated only for the first chunk of each editing round, while subsequent chunks are generated by the frozen backbone. Experimental results demonstrate promising improvements in edit faithfulness and temporal consistency.

Critical strengths
  1. The paper introduces a novel task that extends conventional in-place video editing to continuous editing of future video content.

  2. The proposed Edit-Ignition strategy leverages the long-video generation capabilities of a pretrained backbone without requiring full-model fine-tuning.

Critical weaknesses
  1. Limited Baseline Comparison: The current Helios-Base baseline relies solely on direct prompt replacement, which may fail to effectively capture scene-dependent relative editing instructions. I suggest that the authors include a stronger baseline that employs scene-aware prompt rewriting under the same backbone, history conditioning, and inference settings to better demonstrate the contribution of the proposed editing adapter.

  2. Missing Ablation Studies: The paper lacks ablation studies on several key design choices, including the three attention modules, history corruption, the two-stage training strategy, and anchor resetting. Specifically, the authors should compare two settings: enabling the adapter only for the first text block versus keeping the adapter active throughout the entire editing stage. Such a comparison would help validate the advantages of this core inference strategy in terms of both editing quality and computational efficiency.

  3. Limited Evaluation of Long-Term Editing: Although the paper focuses on infinite video editing, its main quantitative evaluation only considers three consecutive editing rounds. How does the proposed method perform when the number of editing rounds increases significantly (e.g., 10–20 rounds)? I am concerned that long-term editing may lead to error accumulation and forgetting of previously edited attributes.

Questions for authors

Can the proposed method maintain stable performance when handling conflicting or compositional editing instructions?

7 comments
AuthorsOpenReview ↗

Response to Reviewer bNFU (1/2)

We would like to thank you for your careful review and constructive suggestions. Please see our point-by-point responses to all weaknesses (R) and questions (A) below.

[R1] We have added a comparison between our method and Helios-Base with scene-aware prompt rewriting. Following the practice of ChronoEdit, we use Qwen2.5-VL-7B-Instruct to rewrite each relative edit instruction into a scene-aware prompt for Helios-Base. Both Helios-Base variants use the same backbone, history conditioning, and inference settings. We evaluate both variants and our method on the same 200-video benchmark.

MethodEdit faithfulnessVisual qualityPreservationCoherence
Helios-Base + raw edit instruction2.6372.8502.7902.885
Helios-Base + scene-aware prompt rewriting2.9222.9873.1253.180
Ours + raw edit instruction3.8603.7903.9203.930

Scene-aware rewriting improves Helios-Base on all four metrics relative to the raw-instruction reference, but it remains below our method on every metric. This comparison isolates the adapter's contribution from the effect of stronger prompt wording. We have added it to the revised appendix.

[R2-1] Thank you for the detailed suggestion. We have added ablations of the adapter components, history corruption, the two-phase curriculum, anchor resetting, and the number of ignition chunks. We also compare inference cost when the adapter is active for one versus all four chunks in an edit round. All quality results below use the same 200-video benchmark and four VLM metrics as Table 2. Scores range from 1 to 5, with higher values indicating better performance. Given the limited rebuttal period, the variants that require retraining and their full-model control use the same 10,000-step budget for a fair comparison. We will update those results after full retraining.

  1. Adapter components. The full model and both attention variants are trained under the same data budget.
Adapter variantEdit faithfulnessVisual qualityPreservationCoherence
Full model3.7923.7433.8103.860
Without history cross-attention3.6653.6423.2103.510
Without temporal causal self-attention3.6103.5883.7253.745

The modules play distinct roles:

  • Without history cross-attention, preservation drops by 0.600 and coherence by 0.350. History cross-attention links the leading frames of each chunk to the preceding video. Removing it weakens the history anchor, making later chunks less consistent with earlier ones.
  • Without temporal causal self-attention, edit faithfulness drops by 0.182 and visual quality by 0.155. Temporal causal self-attention propagates cues from earlier to later frames within a chunk. Without it, the edit becomes less consistent across frames. Preservation and coherence decline less because history cross-attention still provides effective anchors.
  • Edit cross-attention is the adapter's only explicit path for receiving the edit instruction. Removing it would eliminate the edit signal. We therefore retain it in both trained variants rather than report unconditioned continuation as an editing ablation.
  1. History corruption. We compare 10,000-step checkpoints trained with and without it.
History corruptionEdit faithfulnessVisual qualityPreservationCoherence
Enabled (full model)3.7923.7433.8103.860
Disabled3.7533.6723.6603.680

History corruption improves all four scores. It exposes the adapter to degraded histories during training, helping it handle imperfect frames from its own rollout.

  1. Two-phase curriculum. We compare the Phase 1 checkpoint with its Phase 2 continuation.
Training stageEdit faithfulnessVisual qualityPreservationCoherence
Phase 13.8283.7653.8153.840
Phase 1 + Phase 23.8603.7903.9203.930

Phase 2 improves all four VLM scores. This further supports the role of stage-two training in stabilizing frames far from the history anchor.

  1. Anchor resetting. We apply both policies to the same final checkpoint.
Anchor resetEdit faithfulnessVisual qualityPreservationCoherence
Disabled3.7673.6583.7603.730
Enabled (default)3.8603.7903.9203.930

Disabling anchor resetting lowers all four scores, including preservation by 0.160 and coherence by 0.200. Resetting the anchor to the first edited frame gives subsequent chunks a reference aligned with the current edit. Without this update, the anchor remains tied to the original source and may pull later chunks back toward the pre-edit appearance. The score drops are consistent with anchor resetting's role in keeping the edited stream stable.

AuthorsOpenReview ↗

Response to Reviewer bNFU (2/2)

[R2-2]

  1. Ignition policy and inference cost. Each edit round contains four chunks. We compare one, two, and four ignition chunks using the same final checkpoint. Four ignition chunks keep the adapter active throughout the round.
Ignition chunks per edit roundEdit faithfulnessVisual qualityPreservationCoherence
1 (default)3.8603.7903.9203.930
23.8823.7133.7503.835
4 (adapter throughout)3.7533.5833.1253.275

Two ignition chunks slightly improve faithfulness over one (+0.022), while one chunk gives the highest visual quality, preservation, and coherence. Four ignition chunks keep the adapter active throughout the edit round but reduce generation quality and stability. This decline is consistent with the adapter's training objective. Our adapter is trained to establish a high-quality edit consistent with the preceding video, not to sustain long rollouts within the same edit round. After the first chunk, the backbone's stronger generation prior supports more reliable continuation from the edited history. We therefore use the adapter for one ignition chunk and the backbone for the remaining chunks.

Activating the adapter for fewer chunks also reduces inference cost. On 20 cases, we measure:

Policy (four chunks per round)Median inference time (s/sample) ↓\downarrowEffective edit FPS ↑\uparrow
1 ignition chunk324.691.22
4 ignition chunks825.350.48

The one-chunk policy takes less than half the median inference time and achieves a higher effective edit FPS. Together, these results support our default in both editing quality and efficiency.

[R3] Thank you for raising this concern. We have added ten-round editing videos to the supplementary folder "multiple_round_cases". These videos show that our method applies successive edits beyond the three rounds in the main evaluation. Each requested edit takes effect, and later rounds build on the edited content. We also include a matched Helios-Base video using the same source and edit instructions.

For quantitative comparison, we evaluate 30 IID videos over ten editing rounds with the same VLM protocol for both methods. Scores range from 1 to 5, with higher values indicating better performance.

MetricOursHelios-Base
Edit faithfulness, rounds 1–10 mean4.143.36
Rolling three-round coherence, rounds 3–10 mean3.703.04

Our method remains effective through round 10 and outperforms Helios-Base on both metrics. Some fine details may indeed drift over a long sequence, but edit faithfulness does not collapse across the ten rounds.

[A1] We conducted additional experiments with conflicting and compositional edit instructions. Representative videos are included in the supplementary folders "conflict" and "compositional".

  • Conflicting instructions. We write a single instruction that requests two mutually exclusive changes at the same time, such as moving the camera both up and down or changing the same person into both ladudu and simpsons comic styles. The resulting videos remain fluent, although the two requests cannot both be satisfied. Because no output can fulfill both demands, this experiment examines whether generation remains stable under a contradictory instruction.
  • Compositional instructions. We write a single scene-specific instruction that combines a style change with a camera movement and requests both in the same edit round. The resulting videos show the requested style change alongside the camera movement. This indicates that the adapter can combine these edit types within one instruction.

Therefore, current experiments show our method's stable generation under conflicting instructions and effective joint editing under compositional instructions.

AuthorsOpenReview ↗

Looking Forward to Your Further Feedback

Dear Reviewer bNFU,

We would like to express our sincere thanks once again for your efforts in reviewing our paper. As there are only two days left for discussion, we would kindly inquire if you have any further suggestions or concerns. We believe that our previous responses have adequatetly addressed your concerns. Should you have any further questions, we would greatly appreciate the opportunity to address them promptly.

Best regards,

Authors

Reviewer_sj7fOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces InfinityEdit, a framework for editing ongoing video streams under a sequence of edit instructions. Unlike conventional in-place video editing, which modifies frames within a fixed temporal window, the proposed 'infinite video editing' task requires generating future segments that continues the stream while satisfying new edit requests. The method attaches a lightweight edit adapter to a frozen Helios-Distilled backbone. The adapter contains history cross-attention, temporal causal self-attention, and edit cross-attention. During inference, it is ignited only for the first chunk following a new edit instruction. The paper also introduces a synthetic data construction pipeline and evaluates sequential editing using VBench metrics and VLM-based judgments.

Critical strengths
  1. The manuscript is clearly written and easy to follow.
  2. The paper distinguishes infinite video editing from conventional in-place editing and identifies the need to preserve temporal continuity and previously applied edits as new instructions arrive. This formulation captures a meaningful challenge in interactive and streaming video generation.
Critical weaknesses
  1. The paper provides limited challenging qualitative examples, such as multiple editing rounds. The supplementary material lacks video cases, which is important for assessing whether transitions remain natural when new edit instructions arrive and whether long editing sequences exhibit cumulative artifacts, such as color drift or temporal inconsistencies.
  2. The paper focuses primarily on style transfer and camera-motion editing, leaving more general editing tasks such as object addition, removal, and replacement/swap unexplored. This limited task coverage raises concerns about the generality of the proposed method and may restrict its applicability to broader video-editing scenarios.
  3. The paper lacks several important ablation studies. (1) Adapter component: History Cross-Attention, Temporal Causal Self-Attention, and Edit Cross-Attention should be ablated individually to isolate its contribution. (2) Inference strategy: the proposed ignite-then-continue strategy should be compared against keeping the adapter active in subsequent chunks. (3) Long horizon stability design: history corruption and anchor resetting should be ablated to verify whether these designs effectively mitigate error accumulation and preserve stability across repeated editing rounds.
  4. The paper does not report inference latency or peak GPU memory usage compared with baselines. Such comparisons are necessary to assess the practical efficiency and real-time applicability.
  5. The paper lacks human evaluation of edit faithfulness and stability across editing rounds.
Questions for authors
  1. Could the authors provide more challenging long-horizon editing cases like more editing rounds and varying durations between multiple editing rounds? An analysis of failure cases would also help characterize the method's robustness and identify potential failure modes in long-horizon editing.
  2. Could the method be extended to more general editing tasks beyond style transfer and camera-motion editing, such as object addition, removal, replacement, and even reference-image-guided editing? I am curious whether the authors have evaluated how well the method scales across diverse editing tasks.
  3. Since both the training data and the evaluation benchmark use UltraVideo as the source-video pool, could the authors clarify how they prevent data leakage?
  4. For style editing, the target continuation is generated after applying Qwen-Image-Edit-2511 to the source video's last frame. How do the authors ensure that artifacts or appearance inconsistencies introduced by this preprocessing step do not lead to temporal discontinuities at the source–target boundary?
9 comments
AuthorsOpenReview ↗

Response to Reviewer sj7f (1/3)

We would like to thank you for the detailed reviews and helpful comments. Please see our point-by-point responses to all weaknesses (R) and questions (A) below.

[R1 & W1] Thank you for these suggestions. We have added long-horizon and variable-duration video cases to the supplementary material. We analyze the video examples and failure cases separately below.

  1. Challenging video cases. The supplementary folders "multiple_round_cases" and "varying_duration_cases" contain videos with 10 editing rounds and videos with varying durations across three editing rounds, respectively. In the 10-round examples, later segments continue the edits established in preceding chunks. The variable-duration examples also show coherent scenes within each edit round as its length changes. The adapter establishes the edit in the first chunk, while the base model continues the edited scene in later chunks.

  2. Failure analysis. The attached videos show two concrete limitations:

    • Color drift. In some "american comic" style cases, colors gradually become oversaturated in later chunks after the base model takes over continuation.
    • Revisiting inconsistency. In some sequences with "zoom in" followed by "zoom out", the return view does not fully recover the earlier scene appearance. Our adapter targets only the immediate edit, while the base model has no explicit mechanism to retrieve an earlier view after several generated chunks. These factors both lead to the revisiting inconsistency.

    We attribute these failure cases to two main factors. First, our current pipeline establishes an edit and continues the video, but does not explicitly enforce long-range appearance or viewpoint consistency. Small deviations can therefore enter the conditioning history and accumulate over later chunks. Second, our current training data cannot exhaustively cover all scene and edit combinations, so unfamiliar combinations can still be challenging.

Taken together, the videos show that our method can maintain edits across ten rounds and varying edit durations, while also revealing color drift and revisiting inconsistency in some cases. We plan to broaden the training data to improve robustness across scenes and edit instructions in the future.

[R2 & W2] Thank you for raising this point. Our current results cover the edit types included in our training data, including style, camera, motion, and concept edits. We selected these types by drawing on the semantic controls studied in prior work, particularly Video-As-Prompt [1], and curated scene-grounded training triplets for them. These training triplets largely shape the editing capabilities the model acquires.

We agree that object addition, removal, and replacement are important editing types. Our current training data do not contain dedicated object addition or removal triplets. However, some replacement-like operations overlap with our concept edits. Figure 6 shows our method transforming a person into cartoon-doll and Ladudu appearances across successive rounds while keeping the background continuous. This case provides evidence that our model could extend beyond explicit style and camera control.

Constructing addition and removal data and retraining the model require more time than the rebuttal period allows. We plan to include these cases and broader replacement edits in the next training version.

[1] Bian, Y., Chen, X., Li, Z., Zhi, T., Sang, S., Luo, L., & Xu, Q. (2026, April). Video-as-prompt: Unified semantic control for video generation. In International Conference on Learning Representations (Vol. 2026, pp. 114558-114583).

AuthorsOpenReview ↗

Response to Reviewer sj7f (2/3)

[R3] We have added ablations of the adapter components, ignition policy, history corruption, and anchor resetting on the same 200-video benchmark. Scores range from 1 to 5, with higher values indicating better performance. Given the limited time available during rebuttal, we cannot complete the full training schedule for all the ablations that require retraining. Therefore, to ensure a fair comparison, we train those variants and the full-model control for the same 10,000 steps. We will update these results once full training is complete.

  1. Adapter components. The full model and both attention variants are trained for 10,000 steps under the same data budget.
Adapter variantEdit faithfulnessVisual qualityPreservationCoherence
Full model3.7923.7433.8103.860
Without history cross-attention3.6653.6423.2103.510
Without temporal causal self-attention3.6103.5883.7253.745

The modules play distinct roles:

  • Without history cross-attention, preservation drops by 0.600 and coherence by 0.350. History cross-attention links the leading frames of each chunk to the preceding video. Removing it weakens the history anchor, making later chunks less consistent with earlier ones.
  • Without temporal causal self-attention, edit faithfulness drops by 0.182 and visual quality by 0.155. Temporal causal self-attention propagates cues from earlier to later frames within a chunk. Without it, the edit becomes less consistent across frames. Preservation and coherence decline less because history cross-attention still provides effective anchors.
  • Edit cross-attention is the adapter's only explicit path for receiving the edit instruction. Removing it would eliminate the edit signal. We therefore retain it in both trained variants rather than report unconditioned continuation as an editing ablation.
  1. Ignition policy. In our evaluation setting, each edit round contains four chunks. We compare activating the adapter for one, two, or all four chunks with the same final checkpoint.
Ignition chunks per edit roundEdit faithfulnessVisual qualityPreservationCoherence
1 (default)3.8603.7903.9203.930
23.8823.7133.7503.835
4 (adapter throughout)3.7533.5833.1253.275

Two ignition chunks slightly improve faithfulness over one (+0.022), while one chunk gives the highest visual quality, preservation, and coherence. Four ignition chunks keep the adapter active throughout the edit round but reduce generation quality and stability. This decline is consistent with the adapter's training objective. Our adapter is trained to establish a high-quality edit consistent with the preceding video, not to sustain long rollouts within the same edit round. After the first chunk, the backbone's stronger generation prior supports more reliable continuation from the edited history. We therefore use the adapter for one ignition chunk and the backbone for the remaining chunks.

  1. History corruption. We compare 10,000-step checkpoints trained with and without it.
History corruptionEdit faithfulnessVisual qualityPreservationCoherence
Enabled (full model)3.7923.7433.8103.860
Disabled3.7533.6723.6603.680

History corruption improves all four scores. It exposes the adapter to degraded histories during training, helping it handle imperfect frames from its own rollout.

  1. Anchor resetting. We apply both policies to the same final checkpoint.
Anchor resetEdit faithfulnessVisual qualityPreservationCoherence
Disabled3.7673.6583.7603.730
Enabled (default)3.8603.7903.9203.930

Disabling anchor resetting lowers all four scores, including preservation by 0.160 and coherence by 0.200. Resetting the anchor to the first edited frame gives subsequent chunks a reference aligned with the current edit. Without this update, the anchor remains tied to the original source and may pull later chunks back toward the pre-edit appearance. The score drops are consistent with anchor resetting's role in keeping the edited stream stable.

AuthorsOpenReview ↗

Response to Reviewer sj7f (3/3)

[R4] Thanks for your advice. The following table compares model size, inference latency, and peak GPU memory with the five baselines.

MethodNative resolutionModel size (backbone)Median inference latency (s/sample) ↓\downarrowPeak allocated GPU memory (GiB) ↓\downarrow
Helios-Base384×64014.3B156.8947.12
Lucy-Edit480×8325B742.9432.16
SANA-Streaming704×12802B490.5915.45
Anchor-Forcing480×8321.3B119.2326.33
Infinity-RoPE480×8321.3B75.8622.50
Ours (Phase 1+2)384×64014.3B324.6949.13

We measured 20 matched videos after one warm-up sample. Latency is the median per-sample inference time. Peak memory is the maximum measured per-sample allocation. We also report the native resolutions for each method. Compared with Helios-Base at the same resolution, our method adds 2.01 GiB of peak allocated memory, while it substantially improves edit faithfulness (2.637 to 3.860) and preservation (2.790 to 3.920). Overall, these results show clear gains in editing quality with a modest increase in peak memory over Helios-Base. We plan to improve inference efficiency in future work.

[R5] We have added a human study comparing edit faithfulness and stability across editing rounds with five baselines. For each case, 20 participants watch the edited videos in temporal order and rate them on a five-point scale. We randomize the presentation order of the methods. Cross-edit coherence captures consistency across the editing rounds.

MethodEdit faithfulnessVisual qualityPreservationCoherenceOurs vs. method win rate (%)
Ours3.8283.7653.8153.840—
Helios-Base2.6372.8502.7902.88592.0
Lucy-Edit2.8632.7232.6652.56091.7
SANA-Streaming3.3033.0933.0603.03070.7
Anchor-Forcing2.5452.6201.8303.48098.0
Infinity-RoPE2.6582.6751.8653.26594.7

Our method receives the highest ratings on all four dimensions. The last column shows its pairwise win rate against each baseline, with a majority of wins in every comparison. Together, these results indicate a strong human preference for our method. We have added the study to Appendix E.4 of the revised paper.

[A3] Thanks for raising this question. We prevent data leakage in two ways: we separate training and evaluation source videos, and independently generate scene-specific instructions for each sample and edit round.

  1. Source videos. We use only a subset of UltraVideo source videos for training. The 200 benchmark videos are excluded from the training triplets. No source video appears in both sets.
  2. Edit instructions. For every sample and edit round, we generate a separate instruction using its scene description, selected edit type, and a few sampled frames. This source-specific construction applies to both training and test samples. Benchmark instructions are not copied from the training triplets or reused across other source videos and rounds.

Together, these two measures prevent training-evaluation leakage, even though both sets draw their source videos from UltraVideo.

[A4] Thank you for raising this concern. We screen the synthesized training triplets to reduce preprocessing artifacts and unintended mismatches at the source–target boundary. After Qwen-Image-Edit-2511 produces the edited first frame, we use VLM assessment and manual filtering. We retain examples whose main entities and spatial layout still match the last source frame, apart from the intended edit. This limits the artifacts and inconsistencies that enter the training targets.

For style editing, the requested appearance takes effect at the start of the target continuation. A visible change at this boundary is therefore expected, even when the scene remains spatially consistent. Our current task emphasizes an immediate effective edit and does not explicitly model a gradual visual transition. Some style changes may consequently appear abrupt. In future work, we plan to use a first-and-last-frame video generation model to synthesize a transition between the source's last frame and the edited target's first frame. We would then include this segment in the training target for a version that supports gradual transitions.

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

The paper introduces infinite video editing, a task where an edit instruction is applied to an ongoing video stream, and the edited content must continue faithfully into future segments. The authors identify two challenges, i.e., faithful continuation without aligned source frames and stability under repeated edits. They propose InfinityEdit, a lightweight adapter attached to a frozen streaming generator (Helios-Distilled), comprising history cross-attention, temporal causal self-attention, and edit cross-attention. Training uses flow matching with history corruption and a two-phase curriculum. At inference, an "ignite-then-continue" strategy activates the adapter only on the first chunk after each edit. A data-collection pipeline synthesizes triplets. Experiments on an OOD benchmark show gains in camera-motion accuracy and VLM-judged faithfulness/preservation/coherence.

Critical strengths
  1. This paper distinguishes infinite video editing from in-place editing and prompt switching via Eqs. (1)–(2) and articulates edit-specific challenges, i.e., alignment, continuation, and stability. The adopted architecture to protect the backbone prior is reasonable. Freezing Helios and inserting zero-initialized residual adapters is a coherent design, and the three-module decomposition maps directly onto the three stated challenges.
  2. This paper proposes explicit measures to address ``exposure bias''. History corruption with p=0.9 and the mixture-of-Gaussians sigma sampling are concrete mechanisms tied to the iterative-editing failure mode, and the description of the two-phase curriculum is detailed.
Critical weaknesses
  1. There is no ablation for the three adapter modules or the training recipe. The paper attributes gains to history cross-attention, temporal causal self-attention, and edit cross-attention, and to history corruption and the two-phase curriculum, but no ablation table appears in the main text or appendices. Without module-wise and recipe-wise ablations, the causal attribution of the reported gains is unsupported.
Questions for authors
  1. It would be better to provide module-wise and recipe-wise ablations evaluated on the same benchmark, so the contribution of each design element to Tables 1–2 can be isolated.
  2. How many training triplets survive human filtering, and what is their distribution across the four edit-type groups? Are the benchmark instructions and filtered triplets released alongside the code?
  3. For Table 2, were prompt-switching baselines (†) given any source-frame conditioning at the edit boundary to make Preservation comparable? If not, how should the Preservation margin be interpreted as evidence of method quality rather than task-setting mismatch?
  4. How sensitive are results to the ignition-chunk length and the choice to disable the adapter after the first chunk?
3 comments
AuthorsOpenReview ↗

Response to Reviewer HVuE (1/2)

We would like to thank you for your detailed reviews and insightful questions. Please see our point-by-point responses to all weaknesses (R) and questions (A) below.

[R1 & A1] Thank you for pointing out this gap. We have added separate module-wise and training-recipe ablations evaluated with the four VLM metrics used in Table 2. Scores range from 1 to 5, with higher values indicating better performance. Given the limited time available during rebuttal, we cannot complete the full training schedule for all the ablations that require retraining. Therefore, to ensure a fair comparison, we train those variants and the full-model control for the same 10,000 steps. We will update these results once full training is complete.

Module-wise ablation. We trained the full adapter and both attention variants under the same data budget.

Adapter variantEdit faithfulnessVisual qualityPreservationCoherence
Full model3.7923.7433.8103.860
Without history cross-attention3.6653.6423.2103.510
Without temporal causal self-attention3.6103.5883.7253.745

The modules play distinct roles:

  1. Without history cross-attention, preservation drops by 0.600 and coherence by 0.350. History cross-attention links the leading frames of each chunk to the preceding video. Removing it weakens the history anchor, making later chunks less consistent with earlier ones.
  2. Without temporal causal self-attention, edit faithfulness drops by 0.182 and visual quality by 0.155. Temporal causal self-attention propagates cues from earlier to later frames within a chunk. Without it, the edit becomes less consistent across frames. Preservation and coherence decline less because history cross-attention still provides effective anchors.
  3. Edit cross-attention is the adapter's only explicit path for receiving the edit instruction. Removing it would eliminate the edit signal. We therefore retain it in both trained variants rather than report unconditioned continuation as an editing ablation.

Training-recipe ablation. Our recipe has two key components: history corruption and a two-phase curriculum. We examine them separately.

  1. History corruption. We compare 10,000-step checkpoints trained with and without history corruption.
History corruptionEdit faithfulnessVisual qualityPreservationCoherence
Enabled (full model)3.7923.7433.8103.860
Disabled3.7533.6723.6603.680

History corruption improves all four scores. It exposes the adapter to degraded histories during training, helping it handle imperfect frames from its own rollout.

  1. Two-phase curriculum. We compare the Phase 1 checkpoint with its Phase 2 continuation:
Training stageEdit faithfulnessVisual qualityPreservationCoherence
Phase 13.8283.7653.8153.840
Phase 1 + Phase 23.8603.7903.9203.930

Phase 2 improves all four VLM scores. This further supports the role of stage-two training in stabilizing frames far from the history anchor. We have updated these results in revised Tables 1–2 and Appendix E.2–E.3.1.

[A2] Regarding the data proportion: After filtering, 65,653 training triplets remain. Their distribution across the four edit groups is shown below and has been added to the revised appendix.

Edit groupShare of loaded triplets (%)
Camera29.7
Motion13.7
Style35.9
Concept20.6

Regarding the release of data: Sure. We plan to release the filtered triplets once the paper is accepted.

[A3] Thank you for raising this concern. We clarify the inference setting and Preservation scoring separately.

First, no source frame is supplied to the prompt-switching baselines at edit boundaries. These methods do not support inserting an additional reference frame there. They generate a continuous video while switching prompts, so earlier generated frames can still influence later segments. Other methods condition on their own preceding frames through their native mechanisms. None receives a new source frame as an extra reference at the boundary.

Second, Preservation compares each method's edited segment with its own preceding frames. Each method is scored independently under the same rule, making the cross-method comparison fair. The VLM rates how much content that method preserves across edits, rather than comparing all methods against one shared reference frame.

Thus, the Table 2 Preservation margin reflects each method's ability to maintain continuity within its own edit sequence, rather than alignment to a shared source frame at the edit boundary.

AuthorsOpenReview ↗

Response to Reviewer HVuE (2/2)

[A4] To assess sensitivity, we compare one, two, and four ignition chunks using the same checkpoint. Each edit round contains four chunks, so the four-ignition variant keeps the adapter active throughout the round.

Ignition chunks per edit roundEdit faithfulnessVisual qualityPreservationCoherence
1 (default)3.8603.7903.9203.930
23.8823.7133.7503.835
4 (adapter throughout)3.7533.5833.1253.275

All three settings outperform the baselines in edit faithfulness and visual quality. Two chunks slightly improve faithfulness over one (+0.022), whereas keeping the adapter active for all four chunks reduces generation quality and stability. The four-chunk decline is consistent with the adapter's training objective. The adapter is trained to establish a high-quality edit consistent with the preceding video, not to sustain long rollouts within the same round. For an edit round, the backbone's stronger generation prior supports more reliable continuation in later chunks. We therefore use the adapter for the first chunk and the backbone for the rest, improving quality and stability over long edited videos.

One ignition chunk also improves efficiency because each adapter invocation adds computation. We compare it with adapter activation for all four chunks on 20 cases:

Policy (four chunks per round)Median inference time (s/sample) ↓\downarrowEffective edit FPS ↑\uparrow
1 ignition chunk324.691.22
4 ignition chunks825.350.48

The one-chunk policy takes less than half the median inference time and achieves a higher effective edit FPS. Together, these results motivate our one-chunk default on both performance and efficiency grounds.

AuthorsOpenReview ↗

Thanks for Your Effort

Dear Reviewer HVuE,

Thank you for your positive assessment of our work. We greatly value your perspective and would welcome any suggestions or concerns you may have. We have addressed the concerns raised by the other reviewers in our rebuttal and are incorporating the relevant clarifications and revisions into the manuscript.

As the discussion period comes to a close, we would be grateful for any further feedback you might wish to share. Your input would be valuable as we prepare the final version.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.