acceptodds
Under review as a conference paper at ICLR 2027

EditCast: One-Step Video Editing by Broadcasting an Edited Keyframe

Abstract

Image editors let users refine a frame’s appearance, but carrying edits through a video often requires another costly generation pass. We present EditCast, a one-step broadcaster that turns any image editor into a video editor. Given a source clip, an edited keyframe and its temporal slot, and an instruction, EditCast predicts the full edited clip in a single generator evaluation. We fully fine-tune VACE-1.3B using pixel and perceptual losses on sampled temporal windows. At equal training steps, a matched 30-step flow-matching editor fine-tuned from the same checkpoint performs worse on every keyframe-following and fidelity measure: by 1.7 dB at the keyframe slot, 2.0 dB at the last frame, and 2.2 dB against the reference clip. On KeyEdit-50, our curated 1280 × 720 benchmark with keyframes from four image editors and reference edits, EditCast is over 2 dB closer to the keyframe at the last frame than the strongest propagation baseline. It runs at 6.3 FPS on a single H200, or 16.8 FPS with TinyVAE, making it at least 33 times faster than all 50-step propagation baselines (87 times with TinyVAE). EditCast accepts one or more keyframes at any temporal slot, follows keyframes from different editors, and preserves 93% of the edit on clips 2.5 times the training length.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Editing, Diffusion Model, Temporal Consistency, Distribution Matching Distillation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_iDekOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies keyframe-guided video editing. The user edits one frame with any image editor, and the goal is to carry that edit through the whole clip while keeping the motion of the source video. EditCast fully fine-tunes Wan2.1-VACE-1.3B to predict the edited clip in one generator evaluation. The source clip and the edited keyframe, placed at any temporal slot, go in through VACE's control pathway. The authors train only with pixel and perceptual losses on random temporal windows, and they warp the start noise along the source optical flow. The training pairs come from OpenVE-3M, Ditto-1M and a set of generated H3 pairs. The main claim is that the source and the keyframe almost fix the target, so a one-step model is both faster and more accurate than a matched multi-step editor. On KeyEdit-50 (1280×720), EditCast follows the keyframe more than 2 dB better at the last frame than the strongest propagation baseline. It runs at 6.3 FPS on one H200, about 33× faster than the 50-step propagation baselines, and users prefer it in a blind study. The supplementary videos and HTML page make it easy to check the temporal quality.

Critical strengths
  1. EditCast makes keyframe-guided video editing much faster. It runs at 6.3 FPS on one H200, about 33× faster than the 50-step propagation baselines (Table 1), and a full edit-and-preview round with FLUX.2 takes about 2.5 seconds of GPU time at 480 × 832 (Appendix F). It also scores best on last-frame following, reference fidelity and the VLM judge among the compared methods (Tables 1 and 6), and users picked it first in 48% of trials (Section 4.2).

  2. The experiments show that the keyframe, and not the text, drives the edit. With the source, prompt and seed fixed, 88% of outputs are closest to their own keyframe, against 33% by chance (Table 3). Dropping the text barely changes the result (17.71 vs. 17.86 dB), while dropping the keyframe lowers it to 13.52 dB (Section 4.4). EditCast also keeps the source's dynamic degree, which rules out simply copying the keyframe (Table 5).

  3. The method generalizes beyond its training setup. It handles keyframes at any slot and two keyframes at once, which it never saw in training (Table 8). It keeps 93% of the edit on 81-frame clips, 2.5 times the training length (Section 4.5), and keyframes from two image editors that differ by 11.7 dB give the same edit ratio (Appendix F).

  4. The reporting is transparent and supports reproducibility. The paper shows a failure case under viewpoint change (Figure 9), includes supplementary videos with an HTML viewer, lists the full data pipeline and filtering rules (Appendix A), and plans to release the code, data lists, scoring masks and model weights.

Critical weaknesses
  1. The evidence for better editing quality leans on metrics that are close to the training loss. EditCast trains with pixel and perceptual losses, and Table 1 measures keyframe following with in-mask PSNR against the edited keyframe. At later frames the edited object moves, so a correct output can still differ from the keyframe pixel by pixel. Reference PSNR and LPIPS cover only the 25 KeyEdit-Ref clips (Section 4.1). These scores measure how well the edit is kept. They say little about whether the video stays natural under motion, occlusion or viewpoint change. The VBench scores, the VLM judge, the user study and the keyframe-swap test help here, so this point limits the strength of the quality claim. The paper also lacks a simple baseline that warps the edited keyframe along the source optical flow. Adding this baseline, together with a motion-aware following metric or a video-level reference score, would show how much EditCast gains over simple propagation.

  2. The one-step vs. multi-step comparison changes more than the number of steps. Both models in Table 2 start from the same VACE-1.3B checkpoint and share the conditioning, learning rate and batch size. The one-step model trains with pixel and perceptual losses, while the 30-step model trains with flow matching and runs without classifier-free guidance (Section 4.3). So the gain may come from the loss, the step count, or both. The result also depends on training length. At 2,000 steps the multi-step model has the higher anchor PSNR (30.21 vs. 29.93 dB), and at 8,000 steps the one-step model leads (Table 2). A multi-step model with additional reconstruction losses would help separate the effects of the training objective and the number of inference steps. The same issue affects Section 3.5. The paper credits flow-warped noise for lowering the warping error from 1.89 to 1.81, while Table 4 compares two recipes that differ in several settings.

Questions for authors
  1. The comparison in Table 2 changes both the training objective and the number of inference steps. Could the authors provide a more controlled comparison, such as adding pixel and perceptual losses to the multi-step model, to clarify which factor explains the improvement? It would also be helpful to evaluate the multi-step model with classifier-free guidance.

  2. Could the authors provide a motion-aware keyframe-following metric and an additional video-level quality measure? A simple baseline that propagates the edited keyframe along the source optical flow would also show how much of the improvement comes from the learned model itself.

  3. Table 9 and Appendix E report last-frame following per edit category and source motion. Could the authors add bootstrap intervals over clips for the main tables, and report the reference-based metrics per category? This would show how consistent the gains are on clips with occlusion, viewpoint change or large motion.

  4. Could the authors provide an ablation comparing i.i.d. noise and flow-warped noise while keeping the remaining training settings fixed? Section 3.5 reports a reduction in warping error from 1.89 to 1.81, but Table 4 changes several training settings, making it difficult to isolate the effect of noise initialization.

  5. How were the KeyEdit-Ref reference videos generated? Could the authors clarify whether using reference targets produced by similar generation pipelines to the training data might influence the reference-based fidelity metrics?

Two small points need clarification:

  • Section 4.2 states that 14 raters evaluated 20 clips each, which gives 280 trials, while the paper reports 282.
  • The abstract says the multi-step editor trails "at equal training steps" on every keyframe-following and fidelity measure, but at 2,000 steps it has the higher anchor PSNR (30.21 vs. 29.93 dB, Table 2).
7 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/5)

We thank Reviewer iDek for the thorough review and concrete suggestions, and for the positive assessment of EditCast's speed, generalization, and transparent reporting. We answer each point below.

W1. Evaluation metrics and naturalness

Thank you for this insightful comment. We added a motion-aware following metric, a video-level reference metric, and a direct optical-flow propagation baseline. EditCast leads on both new metrics among the methods scored and outperforms direct flow propagation on reference PSNR, EWARP, and video-level DINOv2 similarity. VBench motion smoothness, VLM visual quality, and the user study also favor EditCast over PropFly.

  • Motion-aware following metric. We align the keyframe to each output frame using source optical flow and compare co-visible pixels. On KeyEdit-Ref, EditCast leads PropFly at the last frame (22.9 vs. 20.7 dB, see Q2).
  • Simple optical-flow propagation baseline. EditCast’s in-region reference PSNR is 6.46 dB higher than that of direct flow propagation over frames 1–32 (24.08 vs. 17.62 dB). EditCast also achieves lower EWARP (1.80 vs. 3.35) and higher video-level DINOv2 similarity (0.990 vs. 0.950, see Q2).
  • Naturalness. Metrics unrelated to the training loss also favor EditCast: VBench motion smoothness 0.993 (PropFly 0.991), VLM visual quality 3.88 (PropFly 3.58), and in the blinded user study 48% of evaluations ranked EditCast first (PropFly 34%).
  • Motion, occlusion, and large image shifts. On KeyEdit-50, EditCast’s last-frame PSNR lead over PropFly is 3.46 dB with source camera motion and 2.00 dB with a static camera. On the 27 DAVIS clips, we detect no additional target appearance drift after reappearance or large image shifts. EditCast retains the edit in seven of eight reappearance events. After frame 32, the 95% intervals for EditCast’s appearance-drift and background PSNR trends include zero. The 30-step editor shows increasing appearance drift (second table below). On selected clips in the updated supplement, EditCast preserves the edited appearance through occlusion and strong camera motion more consistently than several baselines.

Evidence beyond the training loss. We leave MC-PSNR unreported for direct flow propagation because the baseline uses the same warped keyframe that defines the metric’s target. This gives it an advantage by construction. The table compares the baseline on in-region reference PSNR, EWARP, and video-level DINOv2 similarity.

MeasureEditCastPropFlyFlow propagationSource
Motion-aware following MC-PSNR, last frame (KeyEdit-Ref) ↑22.920.7—Q2
Video-level reference similarity, DINOv2 (KeyEdit-Ref) ↑0.9900.9680.950Q2
In-region reference PSNR (KeyEdit-Ref, frames 1–32) ↑24.0820.4817.62Q2
EWARP (KeyEdit-Ref) ↓1.802.383.35Table 1; flow propagation in Q2
VBench motion smoothness (KeyEdit-50) ↑0.9930.991—Table 5
VLM visual quality (KeyEdit-50) ↑3.883.58—Table 6
Ranked first in the blind user study ↑48%34%—Sec. 4.2

Temporal stability on 27 DAVIS clips (first 81 frames). From the DAVIS-2017 annotations, we automatically select clips with occlusion followed by reappearance or large camera moves, together with control clips without that type of event. FLUX.2 [klein] edits frame 0 to produce the keyframes. We measure drift relative to the target’s appearance in the generated clip’s own first frame, where the keyframe is placed, and subtract the corresponding change in the source video. Event effects are differences-in-differences against the corresponding control clips at the same frames. Brackets give 95% bootstrap intervals.

MeasureEditCast30-step editor
Drift after the target reappears+0.003 [−0.031, 0.041]+0.002 [−0.030, 0.037]
Drift after a large image shift−0.008 [−0.019, 0.003]+0.000 [−0.006, 0.008]
Drift trend after frame 32 (per 32 frames)−0.001 [−0.009, 0.007]+0.016 [0.005, 0.029]
Background PSNR change after frame 32 (dB per 32 frames)+0.05 [−0.63, 0.74]−0.60 [−1.26, 0.09]

We will clarify in the revision what each metric measures.

AuthorsOpenReview ↗

Official Comment by Authors (2/5)

W2. What the loss and noise ablations explain

Following the reviewer's suggestion, we separated these factors: the one-step advantage does not come from the loss alone, the multi-step editor's anchor lead at 2k is an early-training effect, and we will correct the attribution for the initial noise.

  • Loss versus step count. Adding the same pixel and LPIPS losses to the 30-step editor improves temporal consistency, but its average last-frame PSNR gain comes mainly from two clips, and most clips barely change. With the same interface, data, and training budget, the one-step recipe still leads this variant by 1.15 dB in last-frame PSNR and 1.75 dB in in-region reference PSNR (see Q1).
  • Results at 2k updates. At 2k, the multi-step editor has slightly higher anchor PSNR (30.21 vs. 29.93 dB), but the one-step model already has higher last-frame PSNR; by 8k, the one-step model leads in anchor PSNR, last-frame PSNR, edit ratio, and EWARP (Table 2). At 2k, training is still in progress and the metrics can fluctuate before convergence, so we base our conclusion on the matched 8k comparison. We will state the 8k setting in the abstract and report the 2k results in Section 4.3.
  • Initial noise. Our wording in Section 3.5 could imply that flow-warped noise alone reduced EWARP from 1.89 to 1.81. That decrease comes from comparing the two training recipes in Table 4. After both models converged, EWARP was 1.856 with i.i.d. noise and 1.802 with flow-warped noise (see Q4). We will correct the wording in Section 3.5.

Q1. Reconstruction losses and classifier-free guidance

Thank you for your suggestions. We ran both. The multi-step model with added reconstruction losses still trails the one-step recipe (by 1.15 dB in last-frame PSNR). Text CFG lowers all three reported PSNR scores, while EWARP decreases slightly from 2.23 to 2.21. The one-step model is also 6.8 times faster than the 30-step editor (same timing protocol as Table 1).

  • Multi-step model with reconstruction losses. The two 30-step models share the data, interface, 8k training updates, and inference setup; the second adds pixel L2L_2 and LPIPS losses on the predicted clean output. With these losses, its EWARP approaches the one-step result, but the one-step recipe still leads by 1.15 dB in last-frame PSNR and 1.75 dB in in-region reference PSNR (table below), so the loss alone does not explain the one-step advantage.
  • CFG. We keep the multi-step checkpoint fixed and add text CFG only at inference. At CFG scales of 2.5 and 4.0, KeyEdit-Ref in-region PSNR drops from 22.29 to 21.97 and 21.63 dB, respectively.

Multi-step editor with reconstruction losses.

ModelLast-frame PSNR, KeyEdit-50 (dB) ↑KeyEdit-Ref in-region PSNR (dB) ↑LPIPS ↓EWARP ↓
30-step, flow matching15.8821.820.1632.28
30-step, flow matching + pixel L2L_2 + LPIPS16.8322.280.1441.90
One-step, matched training17.9824.030.1161.84

Text CFG at inference.

Text CFGAnchor PSNR (KeyEdit-50) ↑Last-frame PSNR (KeyEdit-50) ↑Ref in-region PSNR ↑Ref LPIPS ↓Ref EWARP ↓
Off28.4616.0422.290.1552.23
2.527.9315.9721.970.1762.21
4.027.3915.7821.630.1822.21

We will add the loss ablation to Table 2 and report the CFG comparison in the appendix.

AuthorsOpenReview ↗

Official Comment by Authors (3/5)

Q2. Motion-aware following and flow propagation

Thank you for your suggestions. We added all three. EditCast still leads on the motion-aware and video-level metrics, and simple optical-flow propagation does not reach its results (6.46 dB lower in-region reference PSNR over frames 1–32).

  • Optical-flow propagation baseline. We copy the edited keyframe into frame 0 and warp it to later frames using source optical flow. We use the warped values on co-visible pixels and keep the source elsewhere. Over frames 1–32, EditCast’s in-region reference PSNR is 6.46 dB higher than this baseline’s (24.08 vs. 17.62 dB). EditCast also achieves 3.13 dB higher last-frame PSNR and lower EWARP (1.80 vs. 3.35).
  • Motion-aware following metric (MC-PSNR). We align the keyframe to each output frame with the source optical flow and compare only pixels visible in both. On KeyEdit-Ref, EditCast scores 24.2 dB on average and 22.9 dB at the last frame, both higher than PropFly (22.0 and 20.7 dB). On KeyEdit-Img, mean MC-PSNR over frames 1–32 is 24.9 dB for EditCast and 21.0 dB for PropFly. We leave MC-PSNR unreported for direct flow propagation because it uses the same source-flow warp as the metric’s target, giving the baseline an advantage by construction.
  • Video-level reference metric. We average DINOv2 features over the whole video and compare them with the edited reference video. EditCast scores 0.990, higher than the 30-step editor (0.977), PropFly (0.968), and flow propagation (0.950).
KeyEdit-Ref methodIn-region reference PSNR (dB) ↑LPIPS ↓EWARP ↓Last-frame PSNR (dB) ↑MC-PSNR mean / last (dB) ↑Clip DINOv2 ↑
EditCast24.080.1151.8017.6224.2 / 22.90.990
30-step editor22.070.1582.2315.8323.0 / 21.80.977
PropFly20.480.2062.3815.4722.0 / 20.70.968
Direct flow propagation17.620.2853.3514.49—0.950
NOVA17.530.3292.4014.1817.5 / 16.50.931
AnyV2V18.130.3584.8815.3319.1 / 18.20.918
VDOT11.480.5602.66——0.852
Copy source10.780.5962.0110.16—0.824

In-region reference PSNR and LPIPS are averaged over frames 1–32 for all methods on the 25 KeyEdit-Ref clips. Direct flow propagation copies the edited keyframe exactly into frame 0.

We will add this baseline and both metrics to the revision.

Reviewer_Qud7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a video editing method that propagates the appearance of an edited keyframe throughout a source video. Its main contribution is a one-step prediction of the edited video’s latent representation, enabling fast inference. The method performs well in both the user study and quantitative evaluations. The paper also includes extensive experiments examining keyframe placement, video length, and the relative influence of different conditioning inputs on the editing results.

Critical strengths

A key strength of this work is its one-step generation framework, which enables fast video editing while maintaining good visual quality. The paper also presents extensive experiments examining the effects of keyframe placement, video length, and different conditioning inputs on the editing results. These analyses, together with comprehensive ablation studies and a user study, provide a clearer understanding of the method’s behavior and support its effectiveness in terms of both editing quality and efficiency.

Critical weaknesses

1.The motivation for using intermediate keyframes needs further clarification. Although the paper evaluates different keyframe positions, it remains unclear when editing an intermediate frame offers a meaningful advantage over editing and propagating the first frame. A more targeted comparison with first-frame propagation methods would help establish this benefit. Moreover, while the conditioning experiments show that the edited keyframe has a less influence than the text prompt, they do not fully explain how keyframe selection affects individual objects. For example, does choosing a frame with better object visibility, a different pose, or less occlusion improve object fidelity throughout the output video? more relevant reference as:

[1]FFP-300K: Scaling First-Frame Propagation for Generalizable Video Editing

2.Generalization beyond the Wan backbone remains unverified. The method is implemented using Wan2.1-VACE, and it is unclear whether the same one-step training strategy would work equally well with other video generation architectures. Experiments with another backbone would help establish whether the approach is broadly applicable or depends heavily on the pretrained model and its conditioning interface.

3.Conflicting reference frames are not sufficiently explored. Although the framework supports multiple edited keyframes, these references may specify inconsistent appearances, object attributes, or scene content. It is unclear how the model resolves such conflicts and whether it produces smooth transitions, ignores one reference, or introduces temporal artifacts.

4.The contribution is primarily empirical, with limited theoretical analysis. The paper demonstrates that direct one-step training works well for this task, but provides limited mathematical justification for when and why a single prediction should be sufficient. A more principled discussion of the assumptions underlying this formulation, and the circumstances in which it may fail, would strengthen the contribution.

5.The evaluation of long videos remains limited. The experiments extending beyond the 33-frame training length are useful, but do not establish performance on substantially longer sequences. Further evaluation would help assess whether appearance drift, temporal inconsistencies, or the loss of edited content accumulate over time, particularly under occlusion and large scene changes.

6.The supplementary videos raise concerns about background dynamics. In examples such as 00_editcast.mp4, the background appears to exhibit limited motion. This raises the question of whether the method preserves realistic background dynamics or achieves temporal consistency partly by suppressing motion. More examples with dynamic backgrounds, together with a targeted evaluation of background motion preservation, would help clarify this issue. It is very critical and important question.

Questions for authors

Please address the questions corresponding to the weaknesses listed above. The 6th question is particularly important, as the results shown in the supplementary MP4 videos are not very convincing.

6 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/5)

We thank Reviewer Qud7 for the thoughtful review and for recognizing the fast one-step editing and the extensive experiments. The questions raised are very helpful for improving our manuscript; we answer each point below.

W1. Intermediate keyframes and target visibility

Intermediate keyframes let users edit the frame that suits the edit best: the target may appear only after frame 0, or the pose to edit may occur only mid-clip. Our experiments show that an intermediate keyframe given at its original time position is more accurate than first-frame propagation, and choosing the right frame improves quality further.

  • Direct comparison with first-frame propagation. Moving the keyframe from frame 0 to frame 16 raises EditCast's in-region PSNR against the edited reference from 24.28 to 24.96 dB. Given the same frame-16 keyframe, the first-frame propagation method PropFly, which has no input for the keyframe's time position, drops instead (from 20.64 to 19.08 dB).
  • Which frame to choose. Keeping everything else fixed and changing only the keyframe, choosing the frame where the target is more fully visible raises in-region PSNR by 0.57 dB; at similar visibility, choosing the more representative pose (closer to the clip's other frames) raises it by 1.18 dB. Changing only the random seed moves the score by 0.02 dB, so both differences come from the choice of frame.
  • Keyframe versus text. As Section 4.4 ("Keyframe versus text") reports, the keyframe, not the text, primarily determines the edit: removing the text lowers last-frame PSNR only from 17.86 to 17.71 dB, whereas removing the keyframe lowers it to 13.52 dB.

Keyframe position: first frame vs. intermediate frame. In-region PSNR against the edited reference on KeyEdit-Ref (dB).

MethodKeyframe at frame 0Keyframe at frame 16
EditCast (Table 8)24.2824.96
PropFly20.6419.08

Keyframe selection. We select evaluation pairs from a pool of 1,200 held-out source–edit video pairs and take keyframes from the edited videos. Differences are matched by propagation distance, with paired 95% bootstrap intervals.

ComparisonIn-region PSNR difference (dB) ↑LPIPS difference ↓
Larger − smaller visible target area (area ratio 0.55)+0.57 [0.11, 1.12]−0.010 [−0.023, −0.0005]
More − less representative pose (visible area within 10%)+1.18 [0.54, 1.84]−0.019 [−0.036, −0.006]
Same keyframe, different random seed+0.02 [−0.02, 0.05]+0.0002 [−0.0004, 0.0008]

We will discuss fixed-first-frame methods such as FFP-300K [1] in related work and add the intermediate-frame comparison with PropFly and the keyframe-selection experiment.

References

[1] FFP-300K: Scaling First-Frame Propagation for Generalizable Video Editing. Huang X, Xu C, Luo D, et al. CVPR 2026.

W2. Generalization to other backbones

Thank you for this constructive suggestion. The one-step recipe also works on a second backbone, CogVideoX-5B: with the same backbone, data and budget, the one-step editor follows the keyframe more closely and reconstructs the edited region more faithfully than a 30-step editor, and runs 6.2 times faster (640×320, same timing protocol as Table 1).

  • Setup. CogVideoX has no VACE control pathway, so we feed the source video and the time-aligned edited keyframe by channel concatenation. On this backbone we trained two editors with the same training data, effective batch size of 8 and learning rate: a one-step editor with our recipe, and a 30-step editor with CogVideoX's native v-prediction objective. Both are trained at 640×320 (width × height).
  • Results. At the 640×320 training resolution, on KeyEdit-50 the one-step editor's last-frame PSNR is 2.87 dB higher and its anchor PSNR 7.06 dB higher (better on 47 and 49 of the 50 clips). On KeyEdit-Ref, its in-region PSNR is 5.88 dB higher, and its LPIPS (0.214 vs. 0.436) and warping error (1.73 vs. 1.98) are lower. This matches the one-step versus 30-step result on Wan in Table 2.

One-step and 30-step editors on CogVideoX-5B.

Model (CogVideoX-5B)Last-frame PSNR (KeyEdit-50) ↑Anchor PSNR (KeyEdit-50) ↑Edit ratio (KeyEdit-50) ↑Ref in-region PSNR ↑Ref LPIPS ↓Ref EWARP ↓
One-step17.6722.870.89820.980.2141.73
30-step (v-prediction)14.8015.810.87915.100.4361.98

We will report the full results in the appendix and discuss them in Section 6.

AuthorsOpenReview ↗

Official Comment by Authors (2/5)

W3. Conflicting keyframes

Following your suggestion, we tested appearance and content conflicts. In both settings, EditCast transitions between the two keyframe conditions.

  • Appearance conflict. When the two keyframes specify different appearances for the same target, the two-keyframe output starts closer to the output generated from the first keyframe alone and ends closer to the output generated from the second alone.
  • Content conflict. We use an edited keyframe at one end and the unedited source frame at the other. In both keyframe orders, all 50 test clips end on the expected side of the transition.
  • Temporal consistency. Across the pooled KeyEdit-50 clips, EWARP_src rises from 2.07 to 2.37 in both orderings. The pooled 95% confidence intervals for the increases include zero. EWARP_src can count intended appearance changes as temporal error.

Appearance conflict (KeyEdit-Img): DINO similarity between the two-keyframe output and the two single-keyframe outputs. In the two-keyframe run, A is placed at frame 0 and B at frame 32. A-only and B-only denote the outputs generated using each keyframe alone. Early and late scores average framewise DINO similarity over the first and last eight frames, respectively. The crossover is the frame where the two-keyframe output becomes closer to the B-only output.

Keyframe pairsClipsSimilarity of A and BEarly: A-only / B-onlyLate: A-only / B-onlyCrossover (median frame)
From different editors160.790.964 / 0.8810.854 / 0.98214.5
From the same editor90.910.995 / 0.9780.978 / 0.99613
All250.830.975 / 0.9160.899 / 0.98714

Content conflict on KeyEdit-50 (50 clips): an edited keyframe at one end and the unedited source frame at the other. The edited-pixel fraction is the share of edit-region pixels that clearly differ from the source; EWARP uses source-video optical flow, with paired 95% bootstrap intervals against one keyframe.

Frame 0 / frame 32Edited-pixel fraction (frames 0 / 16 / 32)EWARP_src ↓vs. one keyframe [95% CI]
edited / — (one keyframe)0.75 / 0.64 / 0.622.07—
edited / source0.73 / 0.45 / 0.052.37+0.29 [−0.01, 0.54]
source / edited0.02 / 0.41 / 0.592.37+0.30 [−0.10, 0.57]

We will add the per-frame curves to the appendix.

AuthorsOpenReview ↗

Official Comment by Authors (3/5)

W4. When and why one prediction can suffice

A single prediction can work well when the source video supplies the motion and geometry and the edited keyframe largely determines the target appearance.

  • Example. For a moving-car recoloring edit, the source video supplies the trajectory and the keyframe supplies the color.
  • Squared-error rationale. Let XX denote the inputs (source video, edited keyframes and their time positions, and instruction), YY the target video, and Y^θ\hat Y_\theta the prediction. Let μ=E[Y∣X]\mu=\mathbb{E}[Y\mid X] and mθ=Eϵ[Y^θ∣X]m_\theta=\mathbb{E}_\epsilon[\hat Y_\theta\mid X]. Given XX, the noise ϵ\epsilon is independent of YY. The expected squared pixel error decomposes as

E[∣Y−Y^θ∣2∣X]=E[∣Y−μ∣2∣X]+∣mθ−μ∣2+Eϵ[∣Y^θ−mθ∣2∣X]\mathbb{E}[\\|Y-\hat Y_\theta\\|^2\mid X]=\mathbb{E}[\\|Y-\mu\\|^2\mid X]+\\|m_\theta-\mu\\|^2+\mathbb{E}_\epsilon[\\|\hat Y_\theta-m_\theta\\|^2\mid X]

The terms correspond to target uncertainty, mean prediction error, and variation across seeds.

1. Squared-error optimum. The target variance is fixed. The other two terms are minimized when the prediction equals μ\mu for every seed, yielding the ideal deterministic map X↦μ(X)X\mapsto\mu(X).

2. Role of conditioning. When the source motion and keyframe appearance largely determine the edit, μ\mu can approximate the target well. When multiple appearances are plausible, their conditional mean can differ from any individual target.

3. One-step prediction. Under linear flow matching, xt=(1−t)z0+tϵx_t=(1-t)z_0+t\epsilon. At t=1t=1, x1=ϵx_1=\epsilon is independent of the target given the inputs. The ideal velocity therefore satisfies ϵ−v⋆(ϵ,1,X)=E[z0∣X]\epsilon-v^\star(\epsilon,1,X)=\mathbb{E}[z_0\mid X]. When the target latent is uniquely determined by XX, this ideal one-step update recovers it. EditCast predicts the clean latent at t=1000t=1000 using z^0=ϵ−vθ(ϵ,X)\hat z_0=\epsilon-v_\theta(\epsilon,X) and applies pixel and LPIPS losses to every decoded frame (Secs. 3.3–3.4).

  • Empirical evidence. On KeyEdit-Ref, the PSNR between EditCast outputs from two seeds is 38.1 dB overall and 35.9 dB in the edit region, compared with 26.9 and 24.4 dB for the 30-step editor. In the new matched-data comparison, the one-step model trained for 8,000 updates on the final set of 68,600 video pairs achieves an in-region reference PSNR of 24.03 dB, compared with 21.82 dB for the matched 30-step editor. The 24.00 dB reported in Section 4.3 is for the one-step development model at 8,000 updates.
  • Failure cases. Viewpoint or motion edits depart from the source geometry (Fig. 9). A later keyframe can supply appearance for regions revealed after occlusion. In held-out local-change pairs where the visible target area grows to at least 1.5× its initial size within 33 frames, the frame-0 keyframe alone gives 5.04 dB lower PSNR in newly revealed regions than in initially visible regions. For regions occluded at frame 0 and visible at frame 32, adding a keyframe at frame 32 improves in-region reference PSNR by 6.17 dB on the held-out pairs and 2.37 dB on KeyEdit-Ref. The high-frequency energy ratio also increases on both sets.

Disocclusion: the frame-0 keyframe alone vs. adding a second keyframe at frame 32. We group pixels in the edit region by when they are visible in the source video. We measure in-region PSNR (dB) against the edited reference over frames 1–31. The high-frequency energy ratio is the output’s high-frequency energy divided by that of the edited reference video.

Test setRegion (share of edit region)Keyframe at frame 0 onlyPlus keyframe at frame 32Gain [95% CI]High-frequency energy ratio (one / two keyframes)
Held-out pairsVisible at frame 0 (46%)22.0723.67+1.60 [1.14, 2.07]0.855 / 0.860
Held-out pairsRevealed only by frame 32 (25%)17.0323.20+6.17 [5.20, 7.09]0.791 / 0.843
Held-out pairsVisible in neither (29%)16.9520.37+3.42 [2.67, 4.24]0.857 / 0.879
KeyEdit-RefVisible at frame 0 (74%)24.6825.80+1.12 [0.80, 1.46]0.788 / 0.796
KeyEdit-RefRevealed only by frame 32 (12%)23.1125.47+2.37 [1.84, 2.94]0.762 / 0.799
KeyEdit-RefVisible in neither (14%)22.8423.55+0.71 [0.35, 1.09]0.827 / 0.836

We will add the derivation and seed comparison to the revision and discuss the assumptions, failure cases, and use of extra keyframes in Section 6.

Reviewer_LF2XOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses keyframe following for video editing. Given a source video, an edited keyframe placed at an arbitrary temporal slot, and an edit instruction, the goal is to propagate the keyframe's appearance change throughout the clip while preserving source motion. The authors propose EditCast, a one-step video broadcaster obtained by fully fine-tuning the Wan2.1-VACE-1.3B model through its native visual-control pathway, trained from scratch with pixel and perceptual losses on sampled temporal windows. Additional components include flow-warped start noise and aligned keyframe/source conditioning that supports one or more keyframes at arbitrary slots. Experiments on the proposed KeyEdit-50 benchmark compare against pixel-aligned propagation, first-frame propagation, reference-image propagation, and instruction-only editors, with a keyframe-swap control, a multi-step vs. one-step comparison at equal training steps, and a blind user study.

Critical strengths
  1. Table 2 fine-tunes a 30-step flow-matching editor from the same checkpoint, with the same conditioning interface, learning rate, and effective batch, and even gives the multi-step model advantages. At equal training steps, the one-step model wins on keyframe-slot fidelity, last-frame following, in-region PSNR, and temporal warping, while VBench consistency/aesthetic/imaging stay within 0.004. This substantiates the paper's claim that the multi-step sampler is not a quality ceiling for this task, rather than asserting speed at a quality cost.
  2. In the keyframe-swap experiment, fixing source, instruction, and seed while varying only the keyframe, 88% of 120 outputs are closest by DINO similarity to their own keyframe against a 33% chance level, with a positive margin on all 40 clip-instruction pairs despite high pairwise keyframe similarity. Complementary ablations, dropping the instruction changes last-frame following only from 17.86 to 17.71 dB, while dropping the keyframe drops it to 13.52 dB, support the claim that the keyframe drives the edit, not the text branch.
Critical weaknesses
  1. The main comparison may be confounded by mismatched input interfaces and resolutions, and the manuscript does not fully rule out this explanation. Several baselines receive the keyframe copied directly into frame 0 rather than a slot-aligned keyframe, so their anchor PSNR measures the copied frame itself, and their subsequent-frame generation may operate under a different conditioning regime. AnyV2V is constrained to its released fixed 640×320 resolution, while other methods are letterboxed or resized at 832×480. Although the authors label interface differences per row and repeat the comparison at 832×480 with the same ordering, no experiment gives a strong baseline with the same arbitrary-slot aligned keyframe interface and the same training data, e.g., VACE-1.3B is untrained, and VDOT is a released 4-step model. Consequently, the ">2 dB at the last frame" claim in Sec. 7 cannot yet be attributed unambiguously to the one-step broadcaster itself; it may partly stem from the conditioning interface or the training corpus.
Questions for authors
  1. Can you report a baseline that receives the same arbitrary-slot aligned keyframe conditioning and is trained on the same 68,600 pairs with the same budget? For example, the best checkpoint of the multi-step editor in Table 2, or VACE fine-tuned with a standard multi-step objective, so that the last-frame gain can be attributed to the one-step broadcaster rather than to the conditioning interface or training corpus.
  2. What is the overlap between the editors used to generate KeyEdit-Img keyframes and those used to generate training keyframes? How are the main-table metrics on an established external benchmark (e.g., the OpenVE-Bench clips already used in Sec. 6)?
  3. The multi-step comparison fixes t=1000, 30 Euler steps, shift 3.0, and no CFG. Did the multi-step editor receive the H3 pairs and direction-reversal augmentation of the final recipe, or only the development corpus? Would a stronger multi-step configuration (CFG, more steps, other shifts) change the conclusion in Table 2? Additionly, what happens at 161+ frames or 720p native generation? Does the one-step advantage persist?
  4. Fig.9 shows viewpoint-change failure. What fraction of KeyEdit-50 or OpenVE-Bench clips involve camera motion or viewpoint change, and does the method's advantage over PropFly shrink on that subset? The paper reports the largest lead on fast-motion clips, does this hold when viewpoint changes are separated from object motion?
4 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/4)

We thank Reviewer LF2X for the careful review and detailed questions, and for recognizing the matched multi-step comparison and the keyframe-swap test. We answer each point below.

W1 and Q1. Keyframe conditioning, resolution, and matched multi-step baseline

EditCast retains its last-frame PSNR lead at the same input resolution as AnyV2V. It also leads both 30-step editors when the keyframe interface, training data, and training budget are matched.

  • Keyframe placement. Every method in Table 1 that accepts a keyframe receives the same edited keyframe from source frame 0 through its released interface. In the three † rows, zero-shot VACE and VDOT place the keyframe in frame 0 of the control video and mark it for preservation. NOVA uses it as the first-frame image condition. These three models receive the keyframe as input and reconstruct frame 0, which we score directly. The resulting anchor PSNRs range from 32.54 to 33.74 dB. EditCast generates the anchor jointly with the rest of the clip and achieves a comparable 31.42 dB. At the last frame, EditCast reaches 17.86 dB, leading these baselines by 3.25–7.25 dB. We will correct the paper’s † note in the Table 1 caption and Appendix A to specify that anchor PSNR is measured on the reconstructed output frame.
  • Resolution. For the quality comparison in Table 1, we use the benchmarks’ native 1280×720 inputs, except for AnyV2V, whose released model requires 640×320. EditCast leads in last-frame PSNR, in-region PSNR against the edited reference, LPIPS, EWARP, and VLM score. At matched 640×320, EditCast retains these leads over AnyV2V, including in last-frame PSNR on KeyEdit-50 (17.74 vs. 15.64 dB, Table R1). Appendix Table 7 provides a supplementary comparison at 832×480, close to the training resolution of most baselines. For the 832×480 runs, we align input sizes by resizing or padding as required by each method. EditCast still leads PropFly in last-frame PSNR (17.70 vs. 15.84 dB).
  • Matched 30-step comparison (Q1). We compare the one-step editor with a standard 30-step flow-matching editor and a 30-step variant with added pixel and LPIPS losses. All three use the same slot-aligned keyframe interface and the final training set of 68,600 video pairs, with 8,000 updates and an effective batch size of 8. In Table R2, the one-step editor leads the standard and loss-augmented variants by 2.10 and 1.15 dB in last-frame PSNR, respectively. It also leads both variants in edit ratio, in-region reference PSNR, LPIPS, EWARP, and VLM score. In a separate timing test using the Table 1 protocol, the final EditCast model ran 6.8 times as fast as the matched 30-step flow-matching editor.

Table R1. Comparison at 640×320 input. Edit ratios use the reference edit-area denominator from the Table 1 evaluation.

MethodLast-frame PSNR (KeyEdit-50) ↑Anchor PSNR (KeyEdit-50) ↑Edit ratio (KeyEdit-50) ↑Ref in-region PSNR ↑Ref LPIPS ↓Ref EWARP ↓VLM score (KeyEdit-50) ↑
EditCast17.7427.470.88622.790.2321.984.07
AnyV2V15.6423.380.89618.260.3544.883.63

Table R2. One-step and 30-step editors under matched training. Edit ratios use the reference edit area for this evaluation. This denominator differs from the one in Table R1. We compare edit ratios within each table.

ModelLast-frame PSNR (KeyEdit-50) ↑Anchor PSNR (KeyEdit-50) ↑Edit ratio (KeyEdit-50) ↑Ref in-region PSNR ↑Ref LPIPS ↓Ref EWARP ↓VLM score (KeyEdit-50) ↑
One-step17.9830.710.90924.030.1161.844.08
30-step FM15.8828.490.81321.820.1632.283.96
30-step FM + pixel loss and LPIPS16.8330.340.87122.280.1441.904.01

We will add the contents of both tables to the corresponding parts of the revision.

AuthorsOpenReview ↗

Official Comment by Authors (2/4)

Q2. Editor overlap and external evaluation

Thank you for asking about editor overlap and external evaluation. Only one of the 25 KeyEdit-Img keyframes was generated with an image editor used to construct the training video pairs. EditCast retains its last-frame PSNR lead over PropFly on clips with keyframes from unseen editors and on both external test sets.

  • Editor overlap. During training, we sample keyframes at random from the target clips of the training pairs. The image editors used to construct those pairs were FLUX.1 Kontext-dev for OpenVE-3M, Qwen-Image for Ditto-1M, and FLUX.2 [klein] 9B for H3. Of the 25 KeyEdit-Img keyframes, one uses the same FLUX.1 Kontext-dev model, and nine use FLUX.2 [klein] 4B, a different model in the same family as the 9B model used for H3. The remaining 15 use two editors unseen in training, Gemini 3 Pro Image (8) and Gemini 3.1 Flash Image (7). On these 15 clips, EditCast achieves 16.47 dB in last-frame PSNR, compared with 15.23 dB for PropFly.
  • OpenVE-Bench. The external test sets have no edited reference videos, so we report last-frame keyframe PSNR, edit ratio, EWARP_src, and VLM scores. EWARP_src follows the computation in Table 1 using optical flow from the source video. We use the clips’ native resolution (about 720p) where supported. On OpenVE-Bench, we evaluate 326 clips across six categories. We exclude subtitle edits because they require text rendering, which our training data does not supervise, and discuss camera edits separately in Q4. EditCast leads PropFly by 0.96 dB in last-frame PSNR, with lower EWARP_src (2.02 vs. 2.60) and a higher VLM score (4.01 vs. 3.86, Table R3). The last-frame PSNR lead holds in all six categories and for both landscape and portrait clips. Under the timing protocol in Table 1, EditCast runs about 33 times as fast as PropFly.
  • UNIC ∪ GenProp. On these 50 clips, EditCast achieves the highest last-frame PSNR and lowest EWARP_src. Its VLM score is higher than those of PropFly and the released UNIC outputs (Table R4).

Table R3. OpenVE-Bench: all categories except subtitle and camera edits (326 clips, 52 of them portrait). FLUX.2 [klein] edits frame 0 to provide the keyframe.

MethodVLM ↑Last-frame PSNR ↑Edit ratio ↑EWARP_src ↓
EditCast4.0116.260.742.02
EditCast + TinyVAE3.9916.210.752.32
PropFly-1.3B3.8615.300.802.60
AnyV2V3.6714.490.855.17
NOVA3.5413.710.472.49
VACE-1.3B zero-shot3.0111.780.333.24
VDOT-14B 4-step2.9812.140.362.39
LiveEdit3.0910.460.833.27
VideoCoF-14B3.0011.560.241.60

Table R4. UNIC ∪ GenProp (50 clips).

MethodVLM ↑Last-frame PSNR ↑Edit ratio ↑EWARP_src ↓
EditCast4.1717.180.812.95
EditCast + TinyVAE4.1917.140.823.42
PropFly-1.3B3.8015.630.824.54
NOVA3.6413.880.554.08
AnyV2V3.0912.370.6511.06
VDOT-14B 4-step3.0511.900.373.64
VACE-1.3B zero-shot3.1011.520.334.81
LiveEdit2.519.860.425.50
VideoCoF-14B2.4610.900.193.07
UNIC (released outputs)3.9416.450.945.64

We will add both tables to the appendix.

AuthorsOpenReview ↗

Official Comment by Authors (3/4)

Q3. Multi-step settings, native 720p, and long clips

Thank you for the careful review. The one-step editor retains its last-frame PSNR lead when both models use direction reversal and across the multi-step sampling settings we tested. At native 720p, EditCast also achieves lower warping error on the tested 81-, 97-, and 161-frame clips. Under the timing protocol of Table 1, it runs 6.8 times as fast as the 30-step editor.

  • H3 data and direction reversal. The 30-step editor in Table 2 used the final training set of 68,600 video pairs, including H3, without direction-reversal augmentation. When both editors use this training set with reversal disabled, the one-step editor leads by 2.10 dB in last-frame PSNR (Table R2). For the reversal comparison, both models use the same direction-reversal augmentation and are matched exactly in the number of training samples processed. We trained the 30-step editor with reversal and compared it with a checkpoint from the final one-step recipe. The one-step model still leads by 0.53 dB in last-frame PSNR and 1.90 dB in anchor PSNR, with lower EWARP (see the table below).
  • Multi-step sampling. We keep the Table 2 multi-step checkpoint fixed and vary only the inference settings. Increasing the number of steps to 50 or the shift to 5.0 changes the reported PSNR scores by less than 0.2 dB. Text CFG at 2.5 or 4.0 lowers all three PSNR scores and slightly reduces EWARP from 2.23 to 2.21. The one-step editor retains its last-frame PSNR lead across the settings tested.
  • 720p and longer videos. The results in Table 2 use native 1280×720 generation. On 81- and 97-frame clips, about 2.5 to 3 times the training length, EditCast achieves higher in-region PSNR against the edited reference in every window and lower EWARP than the 30-step editor. For the 161-frame clips, we measure warping error using source-video optical flow because edited reference videos are unavailable. With direct generation, EditCast has lower EWARP_src than the 30-step editor (4.07 vs. 6.64) and retains more than 0.96 of the edit in each of the eight windows after the first. Generating with overlapping 33-frame windows raises EditCast’s anchor PSNR from 18.36 to 29.05 dB and reduces EWARP_src to 2.30.

Direction reversal, 2×2.

SettingLast-frame PSNR (KeyEdit-50) ↑Anchor PSNR (KeyEdit-50) ↑Edit ratio (KeyEdit-50) ↑Ref in-region PSNR ↑Ref LPIPS ↓Ref EWARP ↓VLM score (KeyEdit-50) ↑
One-step, no reversal17.9830.710.90924.030.1161.844.08
30-step FM, no reversal15.8828.490.81321.820.1632.283.96
One-step, final recipe (with reversal)17.7131.100.89823.900.1211.844.07
30-step FM, with reversal17.1829.200.88223.010.1272.114.09

Sampling settings for the Table 2 multi-step editor.

SamplingAnchor PSNR (KeyEdit-50) ↑Last-frame PSNR (KeyEdit-50) ↑Ref in-region PSNR ↑Ref LPIPS ↓Ref EWARP ↓
30 steps, shift 3.0 (Table 2)28.4616.0422.290.1552.23
50 steps, shift 3.028.3516.0022.230.1532.23
30 steps, shift 5.028.4616.0322.330.1562.24
50 steps, shift 5.028.3515.9922.250.1552.23
30 steps + text CFG 2.527.9315.9721.970.1762.21
30 steps + text CFG 4.027.3915.7821.630.1822.21

81 and 97 frames (KeyEdit-Ref, direct generation at 720p).

KeyEdit-Ref, direct at 720pMethodIn-region reference PSNR ↑ (frames 0–32 / 33–48 / 49–80 / 81–96)Last-frame keyframe PSNR ↑Edit ratio ↑EWARP ↓
81 frames, 14 clipsEditCast22.38 / 21.05 / 20.83 / —15.620.931.87
81 frames, 14 clips30-step flow matching19.62 / 18.40 / 18.03 / —13.900.852.79
97 frames, 8 clipsEditCast21.52 / 19.83 / 20.09 / 20.0014.140.961.88
97 frames, 8 clips30-step flow matching18.02 / 16.79 / 16.94 / 16.6111.980.843.12

161 frames. There is no edited reference video; EWARP uses source-video optical flow (the source itself scores 2.82). Edit retention is the fraction of the target that stays edited relative to the first window; the column gives its range over the other eight windows.

MethodAnchor PSNR ↑EWARP_src ↓Edit retention (windows 2–9)VLM score (Qwen3-VL-32B) ↑
EditCast, direct18.364.070.965–0.9904.00
30-step editor, direct15.596.640.962–0.9973.25
EditCast, overlapping 33-frame windows29.052.300.955–0.9804.04

We will add these comparisons to Section 4.3 and the appendix, and discuss the limit of direct generation for very long videos in Section 6.

AuthorsOpenReview ↗

Official Comment by Authors (4/4)

Q4. Source motion and viewpoint-changing edits

EditCast’s last-frame PSNR lead over PropFly persists on clips with source camera motion.

  • Fraction. In KeyEdit-50, the source camera moves noticeably in 14 of the 50 clips (28%). Among the 296 OpenVE-Bench clips of the five spatially aligned edit categories, the camera moves in 78 (26%).
  • Lead under source camera motion. On KeyEdit-50, EditCast leads PropFly by 3.46 dB in last-frame PSNR on clips with camera motion and by 2.00 dB on static-camera clips. On the 78 OpenVE-Bench clips with camera motion, the lead is 0.86 dB, similar to the 0.91 dB lead on static-camera clips. On those 78 clips, EditCast also has lower temporal warping error measured with source-video optical flow (EWARP_src difference: −0.81, 95% CI [−1.11, −0.50]). Splitting KeyEdit-50 by both camera motion and object speed gives four groups. EditCast leads in all four, with a larger margin for faster objects within each camera-motion group.
  • Large camera moves. We also ran EditCast on 27 DAVIS-2017 clips, most with strong background motion. Drift is the distance between the target’s mask-pooled DINO features in the current output frame and in the output’s own first frame, where the keyframe is placed, minus the corresponding change in the source video. In the 14 events where the view moves by at least half a frame width, we detect no additional drift after the move compared with clips without such events at the same frames (−0.008, 95% CI [−0.019, 0.003]). The quantitative comparisons use the full evaluation sets described above. The updated supplement adds selected DAVIS-2017 clips with occlusions or large camera moves and OpenVE-Bench clips with strong camera motion for visual comparison. In these examples, EditCast propagates the keyframe edit more consistently than several baselines while preserving source camera motion. Under the Table 1 timing protocol, EditCast runs about 33 times as fast as PropFly.
  • Edits that request a viewpoint change. OpenVE-Bench has a separate camera-edit category of 43 clips (10% of its 431 clips), outside the five spatially aligned categories above; all of its instructions request a camera viewpoint change. Such edits change the source motion that EditCast is designed to preserve. They are also absent from our training set, which contains only appearance edits (local changes, object addition and removal, background replacement, and global style); none of its instructions requests a viewpoint change. PropFly does better on them (last-frame PSNR 14.17 vs. 10.77 dB). Figure 9 shows exactly this case; we will state this limitation in Section 6.

Camera-motion split: last-frame PSNR of EditCast and PropFly (dB). Camera motion means dominant source motion of at least 0.5 pixels per frame at 640×360. Slow and fast objects are split at the median object motion. Differences are calculated from the displayed means. Brackets give paired 95% bootstrap intervals.

SubsetClipsEditCastPropFlyDifference [95% CI]
KeyEdit-50, static camera3618.0416.04+2.00 [1.18, 2.90]
KeyEdit-50, moving camera1417.3813.92+3.46 [0.81, 6.78]
KeyEdit-50, static × slow objects2017.2515.72+1.53 [0.65, 2.84]
KeyEdit-50, static × fast objects1619.0216.45+2.57 [1.31, 4.02]
KeyEdit-50, moving × slow objects515.6413.29+2.35 [0.17, 6.51]
KeyEdit-50, moving × fast objects918.3514.27+4.08 [0.33, 8.80]
OpenVE aligned edits, static camera21817.4916.58+0.91 [0.69, 1.13]
OpenVE aligned edits, moving camera7813.2012.34+0.86 [0.56, 1.21]
OpenVE camera-edit category (requests a viewpoint change)4310.7714.17−3.40 [−4.35, −2.53]

Comments on the paper

1 comment
AuthorsOpenReview ↗

Invitation to Further Discussion

Dear Reviewers,

We sincerely appreciate the time and effort you have devoted to reviewing our paper. Your comments have helped us improve both the presentation of the method and the experimental analysis. We provided point-by-point responses a few days ago and hope they address your concerns.

As promised in our responses, we have revised the paper; added and corrected text is shown in blue. The key additions are:

  • Matched multi-step comparison (Sec. 4.3, Table 2(b); Reviewers LF2X and iDek). With the same keyframe interface, training data and budget, the one-step editor leads a standard 30-step flow-matching editor by 2.10 dB and a 30-step editor trained with the same pixel and LPIPS losses by 1.15 dB in last-frame PSNR. With direction reversal on both sides, it still leads by 0.53 dB at the last frame and 1.90 dB at the anchor.
  • Second backbone (Sec. 6, Appendix O; Reviewer Qud7). On CogVideoX-5B, the same one-step recipe leads an identically trained 30-step editor by 2.87 dB at the last frame and runs 6.2 times as fast.
  • External benchmarks (Sec. 6, Appendix K; Reviewer LF2X). EditCast has the highest last-frame PSNR on 326 OpenVE-Bench clips and on 50 UNIC ∪ GenProp clips.
  • Long videos, occlusion and background motion (Sec. 4.5, Appendices L and M; Reviewers LF2X and Qud7). EditCast keeps over 0.96 of the edit on 161-frame clips, shows no significant drift after occlusions or large camera moves on 27 DAVIS-2017 clips, and follows the source background motion more closely than PropFly on clips with camera motion.
  • Evaluation (Sec. 4.2, Appendices E and J; Reviewer iDek). EditCast leads on motion-aware and video-level metrics and exceeds an optical-flow propagation baseline by 6.46 dB in in-region PSNR; its gains over PropFly, NOVA and the 30-step editor in last-frame PSNR, in-region PSNR and EWARP have paired bootstrap intervals that exclude zero.
  • When one prediction suffices (Sec. 6, Appendix G; Reviewer Qud7). A squared-error rationale, the agreement of outputs across seeds, and disocclusion, where a second keyframe supplies the hidden content.

The updated supplementary material adds DAVIS-2017 clips with occlusions or large camera moves and OpenVE-Bench clips with strong camera motion.

With only one day left in the discussion period, we would appreciate it if you could review our responses and updates and share any remaining concerns, so that we still have time to address them. Thank you again for your time and consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.