acceptodds
Under review as a conference paper at ICLR 2027

EditStream: A Unified Autoregressive Framework for Interactive Video Generation and Editing

Abstract

Interactive video generation and editing are becoming increasingly important for creative design. In this paper, we introduce EditStream—a unified framework for interactive video generation and editing. EditStream unifies Text-to-Video, Video-to-Video, Editing Propagation, Reference-guided Video Editing, and Camera Pose Change, enabling flexible control over video generation, transformation, and editing within one system. To make the unified model practical for interactive use, we develop a two-stage distillation approach that combines Velocity Moment Matching (VMM) with autoregressive unrolling. VMM matches conditional velocity moments at student-reached intermediate states to preserve generation quality and motion, while unrolling exposes the student to its own autoregressive predictions to improve temporal stability. They alleviate common challenges in few-step autoregressive video generation, including over-saturation, degraded motion, temporal instability, and complex training. EditStream provides a practical and scalable solution that bridges high-quality multi-task diffusion-based video models with interactive creative workflows.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video, GenAI, Real-time

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_4WgSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper unifies five video tasks—T2V, V2V, first-frame edit propagation, reference-guided editing, and camera pose control—in a single autoregressive framework, with a dual-type conditioning interface that is a clean and practical design choice. The main methodological contribution is the VMM-plus-unrolling distillation recipe, which converts a bidirectional teacher to a few-step causal student without offline ODE pairs or teacher causalization, a genuine simplification over prior conversion pipelines. Results show competitive quality and 10.56 FPS throughput at 720p. Weaknesses include the absence of a unified quantitative metric across all five tasks (several are evaluated qualitatively), and no real-time playback at full resolution. The claim of generality across tasks is supported but could be strengthened with a more systematic cross-task ablation. Overall, a useful and well-engineered contribution.

Critical strengths

1.The paper precisely identifies the difficulty of converting a strong multi-task bidirectional video model into a few-step causal student, and convincingly argues that existing conversion pipelines (offline ODE-pair preparation, teacher causalization) are prohibitively costly across multiple tasks.

2.Matching conditional velocity moments in velocity space, combined with autoregressive unrolling, avoids the distribution-matching instability of DMD and the need for offline ODE pairs, aligning supervision more closely with the few-step inference trajectory.

3.It achieves 0.90 s time-to-first-output and 10.56 FPS throughput at 720p on a single H100, and Energy Annealing effectively mitigates over-saturation, demonstrating clear potential for interactive workflows.

Critical weaknesses

1.Lack of a unified quantitative metric across tasks and insufficient ablation. The five tasks are not evaluated under a single quantitative framework; several (e.g., first-frame propagation and egocentric editing) are only compared qualitatively, making it hard to systematically assess the overall benefit of the unified model. Independent contributions of the dual-type conditioning interface, task-mixing ratios, and Energy Annealing hyperparameters are not fully isolated, leaving the causal attribution of gains partly unestablished.

2.Full-resolution playback is at only 10.56 FPS under 720p, the system falls short of playback rate, so the real-time interactive value proposition still highly depends on further inference techniques like decoder distillation or quantization.

3.Long-horizon stability is under-evaluated. Despite supporting streaming generation, error accumulation and cross-chunk consistency over long sequences are only quantitatively characterized to a limited extent, and no systematic comparison against memory/cache-based long-horizon methods is provided.

Questions for authors

What is the performance of the model on long-horizon stability? How to prove the model performs better under all five tasks are united? Others please refer to the weaknesses.

5 comments

Q1:Unified Tasks and Ablations We clarify that our claim is not that a unified model outperforms every single-task specialist. Rather, our conditioning interface enables joint training across tasks, followed by direct conversion into one autoregressive streaming model.

To strengthen quantitative coverage, we add a common assessment of editing propagation and egocentric editing. GPT6.1-Sol evaluates six evenly spaced frames using edit consistency (E), content preservation (P), temporal/motion consistency (T), and visual quality(Q), each in [0,5], with S=8E+5P+4T+3QS=8E+5P+4T+3Q.

Editing propagation (GenPropBench)

MethodE ↑P ↑T ↑Q ↑S ↑
Senorita-2M4.053.953.083.1373.85
AnyV2V2.542.381.621.8544.23
EditStream4.644.644.003.5987.10

Egocentric editing

MethodE ↑P ↑T ↑Q ↑S ↑
EgoEdit4.604.684.133.9088.38
SANA-Streaming2.232.632.582.2547.98
EditStream4.504.533.983.7885.85

On these released EgoEdit examples, EditStream is broadly comparable to EgoEdit and scores substantially above SANA-Streaming, which was not trained on egocentric cases.

For the conditioning interface, matched 2k-step runs give hybrid/full-token EditVerse scores of 6.68/6.64 and RefVIE scores of 2.91/2.95. Hybrid inference is 2.24×/2.36× faster at 480p/720p, while full-token conditioning better preserves backgrounds (2.63 vs. 2.45).

For distillation, paired ablations compare random exit/full rollout within Self-Forcing and VMM/DMD within our recipe, from the same warm-up checkpoint. Full rollout improves VBench from 84.13 to 84.39; VMM/DMD in our recipe obtain 84.42/84.27 and identical EditVerse scores of 7.35, supporting the combined recipe.

For Energy Annealing, Updated Table 5 and Figure 6 compares EA=0.7, no EA, and EA=0.2 over 946 generations. No EA increases color richness but also clipping/flicker; EA=0.2 reduces artifacts with flatter colors, while EA=0.7 balances both. The training-data mixture and sampling ratios are detailed in Appendix A.2.

These evaluations support our goal of unifying multiple generation/editing tasks and efficiently converting the joint model to streaming inference.

Q2: Inference Efficiency Thank you for raising inference efficiency. We benchmark several configurations using student checkpoint and report time to first frame (TTFF) and throughput.

ConfigurationTTFF (s) ↓FPS ↑
Baseline: single GPU, original decoder0.9010.56
TinyVAE decoder only0.8013.0
480p0.7214.5
Async 3-GPU pipeline (encode / DiT / decode)1.2919.8
Async 3-GPU + compiled DiT + TinyVAE decoder0.8222.4

The asynchronous three-GPU pipeline with DiT torch.compile and TinyVAE achieves 22.4 FPS with 0.82 s TTFF, a 2.12× throughput improvement over the single-GPU/original-decoder baseline. Because it uses additional GPUs, this is a system-level throughput gain rather than a per-GPU efficiency gain.

Further kernel optimization, quantization, inference infrastructure improvements, and newer hardware such as B200 could further improve TTFF and FPS. These deployment optimizations are complementary to our method and beyond this paper's scope.

Q3:Long-horizon Stability We originally listed long-horizon generation as a limitation because this paper focuses on the distillation recipe, while long-horizon generation needs further study. Following the reviewer’s suggestion, we now include additional discussion of the inference-pipeline strategies we explored when the generation is beyond the training 5 seconds.

We add a 60 s evaluation of the four-step student (704×1280, 16 FPS). Beyond the 81-frame training length, it retains a 21-latent-frame KV window, 3 static sink frames, and a rolling sink that reintroduces early history. The eviction begins at about 5 s.

We extend 22 EditVerseBench videos (2/edit type) to 961 frames by forward/backward playback. The 160-frame cycle lets us measure drift on repeated source frames. SANA-Streaming uses its released long-streaming checkpoint, default settings, and identical inputs.

MethodCLIP-T last ↑ΔDINO src last ↑ΔSharp. last →1Cyc. PSNR last ↑ΔJumps/min ↓
Ours (static + rolling sink)0.271-0.0020.703-0.0300.9325.44+0.041.18
Ours w/o rolling sink0.248-0.0260.539-0.1950.9817.81-2.762.91
Ours w/o static sink0.266-0.0070.680-0.0541.1624.52+0.281.36
Sliding-window KV0.158-0.1150.071-0.66223.658.88-7.629.76
SANA-Streaming0.245-0.0140.614-0.1531.2514.64-5.925.09

Scores are averaged over videos. “Last”: final 5 s; Δ: last minus first window (Cyc. PSNR: last minus first repeated cycle). CLIP-T: edit/text alignment; DINO src: source structural similarity; Sharp.: sharpness relative to the first window (1=unchanged; ≫1 indicates artifacts); Cyc. PSNR: repeated-frame output consistency; Jumps: motion-compensated discontinuities/min.

Our edits remain stable over 60 s: CLIP-T Δ=-0.002, DINO Δ=-0.030, and Cyc. PSNR=25.44 dB with Δ=+0.04 dB (95% CI includes zero). We observe no measurable cross-cycle drift after eviction on this test.

The rolling sink is the key: removing it lowers Cyc. PSNR to 17.81 dB, versus 24.52 without the static sink. Sliding-window KV collapses. SANA-Streaming's GDN memory shows more drift (Cyc. PSNR Δ=-5.92 dB; DINO Δ=-0.153) and discontinuities (5.09/min).

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

This paper presents EditStream, a unified autoregressive framework for interactive video generation and editing, supporting text/image-to-video generation, video-to-video editing, first-frame editing propagation, reference-guided editing, and camera pose control. The framework adopts a dual-type conditioning mechanism that incorporates pixel-aligned conditions through channel concatenation and non-aligned references through token concatenation. To convert a bidirectional diffusion teacher into an efficient block-causal student, the authors propose a two-stage distillation strategy consisting of Velocity Moment Matching (VMM) warm-up and VMM-Forcing with autoregressive unrolling. An Energy Annealing mechanism is further introduced to mitigate oversaturation caused by classifier-free guidance. The method is evaluated on multiple generation and editing benchmarks, including VBench, EditVerse, RefVIE-Bench, and a camera-control benchmark. Experimental results demonstrate improved motion dynamics and modest gains in overall text-to-video quality compared with Self-Forcing under the same initialization, while maintaining competitive editing performance. The model also achieves a throughput of 10.56 FPS at 720p on a single H100 GPU.

Critical strengths

The paper presents a practical framework that supports multiple video generation and editing tasks within a single autoregressive model. The dual-type conditioning design accommodates both pixel-aligned and non-aligned inputs, enabling flexible control without relying on separate task-specific architectures. This is a meaningful contribution toward unified and efficient video creation, particularly given the broad range of tasks demonstrated and evaluated.

The proposed two-stage distillation strategy provides a practical approach to converting a bidirectional diffusion model into a few-step causal generator without offline ODE-pair preparation or teacher causalization. The experiments show that VMM initialization consistently improves performance over ODE initialization across multiple tasks. Furthermore, VMM-Forcing substantially improves motion dynamics over Self-Forcing under the same initialization, while maintaining competitive editing performance. The reported inference efficiency further supports the practical relevance of the framework.

Critical weaknesses

The paper presents VMM as a key methodological contribution, but the distinction from existing moment-matching approaches is not entirely clear. More importantly, the comparison between Self-Forcing and VMM-Forcing changes both the distillation objective and the rollout strategy. Self-Forcing uses DMD with randomly truncated denoising, whereas VMM-Forcing uses velocity matching with a complete rollout. Without controlling these factors separately, it is difficult to determine what actually drives the reported improvements. This leaves the central methodological claim less convincing than the overall results suggest.

Although the paper presents EditStream as a unified framework for video generation and editing, some tasks, particularly editing propagation and egocentric editing, are evaluated mainly through qualitative examples. In addition, the proposed camera-control metrics rely on heuristic measures of hole filling and image quality, but their correlation with human judgments or actual geometric correctness is not established. These limitations make it difficult to assess how reliably the method generalizes across all supported tasks.

Questions for authors

The comparison between Self-Forcing and VMM-Forcing changes both the distillation objective and the rollout strategy. Can the authors provide a controlled comparison that separates these two factors, using the same initialization and comparable training budgets? Without this comparison, it is difficult to determine whether the gains come from VMM or from completing the full denoising trajectory during training.

The hole-filling and temporal-stability metrics rely on heuristic thresholds and optical-flow errors. Has the authors' evaluation been validated against human judgments or ground-truth target views? In particular, how well do these metrics distinguish plausible hole filling from content that is merely non-black, and how sensitive are the results to the chosen thresholds?

Equation (8) describes the denoising process as starting from the initial noisy state and progressively moving toward the clean endpoint, but Equations (27) and (28) in Appendix A.7.3 appear to describe the reverse order. Which ordering is used in the actual implementation? This discrepancy makes the rollout procedure difficult to reproduce as currently written.

6 comments

Q1:VMM and Controlled Experiments VMM builds on moment matching in the native velocity parameterization of flow-matching models (Sec. 3.2; Appendix A.7). Velocity- and clean-sample-moment matching are theoretically equivalent under an invertible linear parameterization, but clean-sample matching introduces transition-time-dependent gradient scaling. Direct velocity matching removes this implicit scaling and, with inference-aligned sampling, improves convergence.

The original comparison evaluates complete recipes. Directly transferring our learning rates/update ratio to Self-Forcing was unsuccessful. Applying DMD at randomly selected intermediate steps during full rollout also failed due to mismatched assembled-sequence distributions (new appendix). We therefore add paired ablations from the same VMM warm-up checkpoint:

  • A/B: replace random exit with full four-step rollout and final-step loss in Self-Forcing, keeping other settings.
  • C/D: replace VMM with DMD in our full-rollout recipe, keeping other settings.
VariantVBench Quality ↑Semantic ↑Overall ↑EditVerse ↑s/iter. ↓GPU-h ↓vs. C
A: Self-Forcing + random exit85.0380.5284.137.25134.6±7.29571.80×
B: Self-Forcing + full rollout85.3580.5584.397.32149.4±10.010621.99×
C: VMM-forcing + VMM85.3680.6484.427.3574.9±34.75321.00×
D: VMM-forcing + DMD85.3180.1184.277.3572.6±32.45160.97×

A/C are submitted results. Costs use 32 H200 GPUs, global batch 128, accumulation 4. s/iter. is mean±std over the complete update schedule; GPU-h = s/iter. × 32 × 800 / 3600. Timing includes auxiliary/student updates, rollouts, and synchronization, excluding benchmark inference. Cross-recipe comparisons are not necessarily compute-matched.

B improves over A; C remains slightly ahead of B. D is competitive with C, with a small VBench gap favoring VMM and identical editing scores. The gains therefore reflect the combined recipe.

Our implementation reuses rollouts/minibatches for auxiliary and student updates and alternates full-video and first-block updates. Self-Forcing uses separate critic/generator rollouts, while rank-dependent exits can cause synchronization waits. Thus, the measured efficiency gain reflects implementation and scheduling as well as rollout depth.

Q2: EditingPropagation and EgocentricEdit To supplement the qualitative comparisons, we report GPT assessments of editing propagation on GenPropBench and egocentric editing using released EgoEdit examples.

GPT6.1-Sol receives the instruction, source frames, and candidate frames sampled at 0%, 20%, 40%, 60%, 80%, and 100%. It assigns four integer ratings in [0,5]:

  • E: edit consistency
  • P: preservation of unedited region.
  • T: temporal quality.
  • Q: visual quality.

The weighted score prioritizes edit execution: S=8E+5P+4T+3QS=8E+5P+4T+3Q, ranging from 0 to 100.

Editing propagation

MethodE ↑P ↑T ↑Q ↑S ↑
Senorita-2M4.053.953.083.1373.85
AnyV2V2.542.381.621.8544.23
EditStream4.644.644.003.5987.10

EditStream scores 87.10, versus 73.85 for Senorita-2M and 44.23 for AnyV2V, with higher mean ratings on all four dimensions.

Egocentric editing

MethodE ↑P ↑T ↑Q ↑S ↑
EgoEdit4.604.684.133.9088.38
SANA-Streaming2.232.632.582.2547.98
EditStream4.504.533.983.7885.85

On this subset of released EgoEdit examples, EditStream scores slightly below EgoEdit but remains broadly comparable, while substantially outperforming SANA-Streaming, which was not trained on egocentric cases.

Q3:Camera-control Metrics Our benchmark contains monocular source videos, without ground-truth videos for the target camera trajectories; target-view PSNR/LPIPS therefore cannot be computed. VBench SC/BC/TF/MS measure overall quality and temporal consistency. The reprojection metrics target a specific failure: unfilled disocclusions.

HC measures the fraction of hole pixels that are no longer black. GHF combines BRISQUE-based hole quality with visible-region SSIM. These metrics can only diagnose missing or degraded filling rather than semantic or geometric correctness.

Across 472 entries in our benchmark, on average holes occupy 14.3% of a frame. We replace only the hole regions in 24 scenes (3 frames each) with controlled degradations:

  • HC: gray/noise both score 100.0, versus 96.9 for generated outputs, confirming that HC measures occupancy.
  • GHF hole quality: decreases from 59.6 to 56.1/34.7/31.8/12.3 when 5%/10%/20%/50% of the hole is blackened, reaching 0 for fully black holes.
  • Generated outputs rank above gray and noise in 21/24 scenes (59.6 vs. 51.6/46.0). Blur (60.4) and misaligned copied texture (60.1) are not penalized, because BRISQUE measures natural-image statistics rather than contextual consistency.

For threshold sensitivity, we re-score all 472 entries and eight methods under 16 settings: black threshold {5,10,15,20}/255 × fill gate {0.80,0.85,0.90,0.95}. Absolute GHF values shift by 7–20 points, mainly at the strictest gate (0.95) and threshold (20), but the default ranking is preserved in 12/16 settings. Pairwise reversals occur only under the strictest settings; one involves TrajCrafter and our teacher, separated by just 0.2 points by default. All pairs outside these reversals retain their order. For HC, V+V remains highest and O+S lowest at every threshold.

While checking optical-flow errors, we found that HTS applied forward flow in the wrong warping direction, making the residual sensitive to motion magnitude rather than flicker. We decide to remove HTS because of this implementation and robustness issue. Whole-video temporal consistency measurement remains covered by VBench TF/MS.

Reviewer_4ZxUOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a relatively comprehensive unified framework for streaming generation and editing, but the main issue is that the individual benefits of each design element are not yet clear: the optimal configurations in the distillation comparisons do not fully align with the computational budget, making it difficult to determine how much of the improvement stems from the VMM itself; the hybrid conditional interface also lacks direct ablation studies. Furthermore, “interactivity” is currently manifested primarily through low-latency output, and there is still a lack of sufficient validation regarding the responsiveness and continuity when switching instructions or reference images during the generation process.

Critical strengths
  • The system is comprehensive and practical. A single model supports generation, video editing, reference image guidance, and shot control, while enabling low-latency block-wise output that balances task coverage and inference efficiency.
  • VMM initialization yields clear results and features a relatively straightforward process. There is no need to pre-generate ODE training pairs or train additional causal teachers; while keeping the subsequent Self-Forcing process unchanged, VMM initialization leads to improvements across multiple tasks.
Critical weaknesses
  • The comparison of distillation methods is not entirely fair. Although control experiments with identical initializations were provided, the learning rates, auxiliary model update ratios, and rollout computational costs still vary across different approaches. Without a comparison under uniform configurations and similar computational budgets, it is difficult to determine how much of the benefit stems from the VMM objective itself.
  • The independent contribution of the unified conditional interface has not been verified. The paper does not adequately compare the quality, GPU memory usage, and speed of the hybrid channel/token concatenation method with other conditional injection methods; therefore, while it primarily demonstrates the feasibility of the entire system, it has not yet clearly elucidated the extent of the additional benefits derived from this interface design.
Questions for authors

I look forward to seeing further results from these experiments. If time is tight, you can choose to simplify the setup for some of the experiments.

  • Starting from the same warm-up checkpoint, with fixed data and optimization settings, we compare three groups: “DMD + random early stopping,” “DMD + full four-step rollout,” and “VMM + full four-step rollout.” Each group is trained for 800 steps; we report the quality of generated and edited text, as well as GPU-hours, and supplement the analysis with comparisons under the same computational budget to determine the respective contributions of the rollout and VMM losses.

  • Conditional Interface Ablation: Starting from the same Wan initialization, using the same data and number of training steps, we compare the paper’s hybrid channel/token concatenation with full token concatenation, ensuring both groups receive the same conditional information. We report editing quality, content retention rate, peak VRAM usage, and training/inference speed at 480p and 720p.

  • You may select 40 videos to test switching between text instructions and reference images. Switch from condition A to B after the third video block is outputted, and include control groups that always use A or always use B. Report the response latency for the switch, the editing success rate within the two blocks following the switch, and the continuity of unedited regions; the time required for cache recalculation or history reset is also included in the latency.

5 comments

Q1: Distillation Comparison

Thank you for this suggestion. The original comparison evaluates complete Self-Forcing and VMM-forcing recipes, which differ in rollout, first-block updates, loss placement, auxiliary/student update ratio, learning rates, and also the implementation. We tried to keep Self-Forcing's original optimization settings on purpose to make the results more faithful.

Following the reviewer's suggestion, two initial attempts were unsuccessful: (1) transferring our learning rates and update ratio of VMM-Forcing recipe to Self-Forcing recipe, indicating recipe sensitivity to hyper-parameters; (2) using full rollout in self-Forcing recipe while applying DMD supervision at randomly selected intermediate steps, where the assembled sequence distribution was mismatched with the bidirectional teacher (We explain the reasons using visualization in the updated appendix).

We therefore conduct paired ablations:

  • A/B (ablating rollout): using Self-Forcing recipe and replacing Self-Forcing's random exit with full four-step rollout, applying the loss only at the final step and retaining other settings.
  • C/D (ablating objective): using the VMM-Forcing full-rollout recipe and replacing VMM with DMD loss, retaining other settings.

All start from the same VMM warm-up checkpoint; A/C are the submitted results in the paper.

VariantVBench Quality ↑Semantic ↑Overall ↑EditVerse ↑s/iter. ↓GPU-h ↓vs. C
A: Self-Forcing + random exit85.0380.5284.137.25134.6±7.29571.80×
B: Self-Forcing + full rollout85.3580.5584.397.32149.4±10.010621.99×
C: VMM-forcing + VMM85.3680.6484.427.3574.9±34.75321.00×
D: VMM-forcing + DMD85.3180.1184.277.3572.6±32.45160.97×

Efficiency is measured on 32 H200 GPUs (4 nodes), global batch 128, gradient accumulation 4. s/iter. is mean±std over the complete update schedule, including auxiliary/student updates, rollouts, and synchronization. GPU-h = s/iter. × 32 × 800 / 3600; vs. C (final recipe) is relative cost.

B improves over A, supporting full rollout within Self-Forcing has great benefits; C remains slightly ahead of B though. D is competitive with C, with a small VBench gap favoring VMM and identical EditVerse scores. The objectives are broadly comparable in this setting.

To explain the efficiency gap, our recipe implementation reuses rollouts/minibatches across auxiliary and student updates and alternates full-video with first-block updates. While self-Forcing recipe generates separate critic/generator rollouts; rank-dependent exit depths can also make shorter rollouts wait for longer ones, slowing the training. These differences explain why fewer rollout steps need not reduce measured cost.

Therefore, the efficiency advantage and performance gains belong to the complete implementation and training recipe of VMM-Forcing recipe.

Q2: Conditional Interface Our hybrid interface originally targets low latency. We compare it with full token concatenation using the same Wan initialization, conditions, data, optimization, and 2,000 iterations (sorry, limited by rebuttal time and compute). For 81 frames, full token concatenation doubles the DiT sequence length: 8,190→16,380 at 480p and 18,480→36,960 at 720p.

InterfaceTrainingEditVerse quality ↑RefVIE All ↑BG ↑FG ↑
Hybrid2k iterations6.682.912.453.15
Full token2k iterations6.642.952.633.11

At equal steps, the two interfaces achieve comparable editing quality, while full token does better in large-region editing like background reference editing.

Efficiency is measured on a fixed 81-frame source/target pair using H100 80GB GPUs:

ResolutionConditioningTrain s/step ↓Train DiT mem ↓Infer s/video ↓Infer DiT mem ↓
480pHybrid3.113.427.80.7
480pFull token4.396.962.32.7
720pHybrid6.998.079.61.6
720pFull token11.5517.5188.16.1

Memory is in GiB; training memory is per GPU. DiT mem is additional memory above resident weights, optimizer states, T5, and VAE. Resolutions are 480×832 and 704×1280.

Full token concatenation increases training time by 1.41×/1.65× and DiT training memory by 2.0×/2.2× at 480p/720p; inference latency increases by 2.24×/2.36× and DiT memory by approximately 3.8×. Quadratic attention cost contributes to the latency gap.

These results show a quality–efficiency trade-off: full token concatenation better preserves backgrounds, while hybrid conditioning achieves comparable editing quality at equal steps with 2.2–2.4× faster inference. We adopt hybrid conditioning for streaming latency and memory requirements and will add this comparison and trade-off to the paper. We will also leaving the study of more advanced interface for future work.

Q3:Instruction Switching We first clarify interaction in our setting. T2V interaction typically generates new content and motion on the fly. In V2V, subject and camera motion come from the source video or webcam; we refer interaction as the one between the real person and the inserted objects and background. Our default handles a new instruction by fully resetting the cache and restarting at the switch point, without additional inference-time continuity mechanisms.

Following the suggestion, we test a demanding case on 40 EditVerse videos: A and B differ in one attribute, e.g., black glasses frames/dark lenses → red frames/red-tinted lenses. The object, placement, and other descriptions remain fixed, so the desired transition changes only that attribute while preserving others in the A editing. This tests transitions between edits. However, our model was not trained for editing interpolation, this test serves as a stress test rather than our intended interaction setting.

SettingLatency (ms)Success@2 blocks ↑Boundary residual ↓PSNR (unedited) ↑
Always-A—10.0%0.005133.00
Always-B—90.0%0.005333.01
Keep1040±815.0%0.005132.96
Recompute1348±1330.0%0.010131.72
Reset (ours, default)1150±2192.5%0.021831.39

A→B occurs after the third block (pixel frame 33). Latency includes text encoding, cache reset/recomputation, denoising, and VAE decoding of the first B block, on a single H100 with bf16 and no acceleration.

Our default reset follows B within 1.15 s, approximately 160 ms beyond a regular block, and achieves success comparable to Always-B (92.5% vs. 90.0%). The discontinuity is localized to the boundary frame; afterwards, the unedited-region residual returns near the control level (0.0054 vs. 0.0045), with a 1.6 dB reduction in unedited PSNR. Keeping (no change to the cache) or recomputing the cache better preserves continuity but remains more anchored to A (15.0%/30.0% success). That's because inside the cache, the history still contains the context of A.

Smooth, in-place attribute transitions could potentially benefit from explicit interpolation training or inference-time re-caching, which we regard as promising future work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.