acceptodds
Under review as a conference paper at ICLR 2027

EverStyle: Long-Horizon Consistent Streaming Video Stylization

Abstract

Streaming video stylization continuously transforms input video frames into user-specified visual styles, supporting real-time applications such as live broadcasting and interactive content creation. Recent diffusion-based methods have made substantial progress toward efficient and temporally coherent streaming stylization, yet maintaining consistent stylization throughout an open-ended rollout remains challenging. In particular, previously seen content is usually stylized differently when it reappears after leaving the active context, despite smooth local segments. To address these, we propose **EverStyle**, a novel long-horizon consistent streaming video stylization framework. To recover the established appearance of recurring content, we design Source-Indexed Stylization Memory (SISM), which stores and retrieves source-indexed stylization residuals, and Replay-Based Stylization Consistency Regularization (RSCR), which trains the generator to reproduce earlier stylization under different temporal contexts. Yet reproducing earlier appearance alone, without constraining new content to the same distribution, cannot keep the evolving stylization temporally consistent, leading to global drift. To suppress this, we introduce Early-Anchored Stylization Distribution Matching (EA-SDM), which jointly evaluates early and current generations to align with the same style without content correspondence. Experiments demonstrate that EverStyle maintains competitive stylization quality, while substantially improving long-horizon consistency under both natural scene revisitation and exact content recurrence settings, sustaining consistent stylization over sequences exceeding 130,000 frames.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Streaming Video Stylization, Long-Horizon Consistency

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_i22SOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses long-horizon style inconsistency in open-ended streaming video stylization. Existing diffusion-based methods generate locally smooth frames but produce mismatched styles when objects or scenes reappear after leaving the active context, leading to cumulative global style drift. The authors propose EverStyle with three core modules: SISM stores and retrieves stylization residuals for recurring content; RSCR trains the generator to reproduce past stylization under different temporal contexts; EA-SDM aligns outputs to the initial style distribution to reduce global drift. Experiments show EverStyle preserves stylization quality and improves long-horizon consistency for revisited scenes and recurring content. Ablation studies verify the effectiveness of each main module.

Critical strengths
  1. This work targets a practical problem unique to open-ended streaming stylization. When content disappears from the active context and later reappears, its stylized appearance often changes. Existing temporal consistency methods mainly reduce short-term flicker and overlook this long-horizon recurrence issue. The proposed memory and distribution anchoring designs target this open-ended streaming setting.
  2. The framework achieves strong qualitative and quantitative results against baselines while maintaining competitive stylization quality. Ablation studies confirm the individual contribution of SISM, RSCR and EA-SDM. Tests on sequences over 130k frames demonstrate stable long-streaming operation.
Critical weaknesses
  1. The central claim of improved consistency for recurring objects and reduced global style drift over long sequences lacks dedicated quantitative experiments. Visual results offer partial evidence, but there are no tailored metrics or controlled tests to measure appearance mismatch of reappeared content or the speed of cumulative drift.
  2. The paper has no failure case visualization or analysis. The performance boundary of EverStyle remains unclear, such as cases with residual style drift or failed SISM retrieval.
  3. The presentation leans too heavily on technical details, making the motivation less intuitive. The method section would benefit from a high-level technical overview. In addition, the overview figure Fig.2 appears cluttered and hard to follow. Excessive fine-grained details obscure the main pipeline and weaken readability.
Questions for authors
  1. Could you add dedicated quantitative metrics to measure appearance mismatch for recurring objects and the rate of cumulative global style drift, to validate your core claim about long-horizon consistency?
  2. What are EverStyle’s primary failure modes? Please describe scenarios where SISM retrieval fails or EA-SDM cannot fully stop long-horizon style drift, and explain their underlying causes?
  3. Could you add a high-level technical overview to clarify the core pipeline?
5 comments
AuthorsOpenReview ↗

Author Response (1/n)

Q1: Evaluations about long-horizon consistency.

A1: Thank you for highlighting the need for more direct quantitative evidence. We clarify that Sec. 4.2 (Table 1) already provides two controlled quantitative evaluations of recurring-content consistency. IVEBench-Cycle compares identical source frames across distant cycles, while CAM evaluates revisits selected using field-of-view overlap annotations. Both quantify the appearance mismatch using LPIPS/DINO.

Following your suggestion, we further add a dedicated style metric, Contrastive Style Descriptors (CSD), for matched-frame comparisons, and measure cumulative drift by tracking the cross-frame CSD at the 25%, 50%, 75%, and final cycle positions. EverStyle achieves the highest matched-frame CSD and maintains the highest cross-frame CSD at all four positions, with a net change of +0.0009+0.0009 from the 25% position to the final cycle. A4 provides the protocols and full results.

We hope these additions can help resolve your doubt and strengthen the quantitative validation of recurring-content consistency and long-horizon style retention.

Q2: More discussions of failures.

A2: Thank you for this important comment. Following your suggestion, we have added failure case visualizations and the corresponding analysis to Appx. B.11 and Fig. 16 (highlighted in blue).

For residual style drift, Fig. 16b shows an oil-painting result whose broad palette and distant ridgeline remain stable, while fine brushstrokes and rock contours change between frames 773 and 5,923. For persistent errors, Fig. 16a shows a case where a dim interior is rendered as bright storefronts and the error persists across recurrences: since SISM verifies agreement between observations, an error shared by both can pass verification and be reused.

For failed SISM retrieval, the typical scenarios are large viewpoint or appearance changes and the eviction of a useful entry. When no reliable correspondence remains, confidence calibration suppresses the correction and generation falls back to the original attention path.

These cases clarify the performance boundary: EverStyle works best when recurring content keeps a reliable source correspondence under a fixed target style. A5 further explains the underlying causes. We hope these examples and clarifications can address your concern.

Q3: Presentation optimization.

A3: Thank you for this helpful suggestion. Following your recommendation, we have added a high-level technical overview and a simplified pipeline figure to make the motivation and the main pipeline easier to follow (highlighted in blue). Due to the page limit, they are placed in Appx. A.1 and referenced from the caption of Fig. 2.

The overview explains the two requirements behind our design: recovering the established appearance of recurring content and maintaining the overall style as content changes. The accompanying Fig. 9 shows one source chunk passing through the streaming generator, with SISM represented as a single side branch in panel (a), and RSCR and EA-SDM as training objectives in panel (b). Layer-level details remain in Fig. 2 and Appx. A.2.

We hope these additions can make the motivation and the core pipeline of our method clearer.

AuthorsOpenReview ↗

Author Response (2/2)

Q4: Validation about long-horizon consistency.

A4: Thank you for this valuable suggestion. Following your recommendation, we added dedicated quantitative metrics for both aspects on IVEBench-Cycle: (1) the appearance mismatch of reappearing content and (2) the cumulative global style drift over time. In short, EverStyle has the lowest mismatch among all five methods, and its style similarity shows no decline over time (a net change of +0.0009+0.0009).

  • Appearance mismatch of reappearing content. IVEBench-Cycle repeats source videos, so identical content reappears at known frame positions. We compare corresponding frames from the first and final complete cycles using LPIPS/DINO and Contrastive Style Descriptors (CSD), a metric specifically designed for image style similarity (Somepalli et al., Measuring Style Similarity in Diffusion Models, ECCV 2024). The results are as follows:

    MethodLPIPS ↓\downarrowDINO ↑\uparrowMatched-frame CSD ↑\uparrow
    StreamV2V0.04470.97620.9665
    SANA-Streaming0.42920.72990.7138
    LiveEdit0.14440.90130.9184
    JoyAI-Video-Edit0.08570.96490.8943
    EverStyle0.04360.97780.9802

    From these results, we observe that EverStyle achieves the best scores on all three metrics, supporting consistent reproduction of the earlier stylized appearance.

  • Cumulative global style drift. We use the first complete cycle as a fixed reference and track the cross-frame CSD at the 25%, 50%, 75%, and final-cycle positions, comparing temporally separated positions to reduce reliance on matching frame content. The results are as follows:

    Method25% CSD ↑\uparrow50% CSD ↑\uparrow75% CSD ↑\uparrow100% CSD ↑\uparrowΔ\Delta CSD (25% →\to 100%)
    StreamV2V0.75220.75120.75050.7491−0.0031-0.0031
    SANA-Streaming0.67950.69330.68100.6523−0.0272-0.0272
    LiveEdit0.80680.80880.81050.8069+0.0001+0.0001
    JoyAI-Video-Edit0.79390.79160.79470.7936−0.0003-0.0003
    EverStyle0.87270.87060.87130.8736+0.0009+0.0009

    From these results, we observe that EverStyle maintains the highest CSD at every position, with a net change of only +0.0009+0.0009 from the 25% position to the final cycle, i.e., its style similarity does not decline over time.

We have added these evaluations to Appx. B.3 (highlighted in blue). We hope these additional metrics can help resolve your doubt and strengthen the quantitative validation of our core claim about long-horizon consistency.

Q5: More discussions and explanations regarding failure modes.

A5: Thank you for raising this question. The primary failure modes are unreliable historical corrections and local drift despite a stable overall style. We describe the scenarios and their underlying causes below.

  • Scenarios where SISM retrieval fails. Large viewpoint or appearance changes can weaken the source-feature correspondence, and eviction from the fixed-size bank can remove a useful historical residual. When no reliable match remains, confidence calibration suppresses the correction and generation falls back to the original sink-and-recent path. In addition, verification can accept an incorrect stylization shared by two observations, allowing the error to persist across revisits, as shown in Fig. 16a.
  • Scenarios where EA-SDM cannot fully stop long-horizon style drift. EA-SDM constrains early and current frames jointly at the distribution level without pixel-wise correspondence. It can therefore preserve the overall style while fine brushstrokes or contours change, as in the oil-painting example in Fig. 16b.

We have added these failure cases and the corresponding discussion to Appx. B.11 (highlighted in blue). We hope these explanations can address your question and make the causes of the remaining failures clearer.

Q6: Adding high-level overview.

A6: Thank you for this suggestion. Following your recommendation, we have added a concise high-level overview and a simplified pipeline figure (Appx. A.1 and Fig. 9, highlighted in blue), and the caption of Fig. 2 now points to them. The overview states the two consistency requirements, i.e., recovering the earlier stylization of recurring content and maintaining the overall style as content changes, and how SISM, RSCR, and EA-SDM address them. The figure separates streaming generation from the training objectives and omits layer-level details, as a companion to Fig. 2. We hope this addition can make the core pipeline easier to follow.

AuthorsOpenReview ↗

Thanks for your comments, which helped us further improve the analysis and presentation

Dear Reviewer i22S,

Thank you again for your time and thoughtful comments during the review process. Inspired by your comments, the revised paper now includes more quantitative results on long-horizon consistency, an analysis of failure cases, and a high-level overview of the method. If you have any further questions, we are happy to discuss and address them.

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

EverStyle augments a LongLive-style two-step causal video stylizer with source-indexed stylization memory (SISM), replay-based consistency regularization (RSCR), and early-anchored stylization distribution matching (EA-SDM). SISM stores full-dimensional stylization residuals under compressed source addresses in a verified, diversity-retained fixed bank of 15 frames per attention layer. RSCR trains reproduction of historical stylizations under new rollout contexts; EA-SDM places early detached outputs alongside current outputs in the DMD clip. The model is fine-tuned/distilled from Wan2.2-5B with Ditto-1M style data. Experiments cover 143 IVEBench videos, ten looped long sequences, five natural-revisit sequences, an ultra-long qualitative rollout, component ablations, and a user study.

Critical strengths

S1:The problem decomposition is useful and the mechanisms address distinct failure modes. Source-relative residual memory separates reusable style choices from source geometry, replay supplies changed-context supervision, and early anchoring constrains current generation without requiring recurrence. Per-layer verification and diversity retention provide a concrete bounded-state design rather than an unspecified growing cache. S2:The reported long-revisit improvements are substantial and supported by within-model ablations. CAM LPIPS is 0.2830 versus the best competing 0.4224; removing SISM raises it to 0.3977, removing RSCR to 0.3299, and removing EA-SDM to 0.3416. IVEBench Total Score is 0.758/0.756, and the qualitative long rollout remains visibly stylized. These results support a meaningful consistency contribution beyond a short-clip quality comparison.

Critical weaknesses

W1:A bounded memory state is demonstrated by design, but the practical streaming/real-time claim is not measured. Neither the main paper, appendix nor supplementary webpage text reports hardware-specific FPS, first-frame/chunk latency or peak memory. All-layer nearest-neighbour retrieval and an additional attention branch can have nontrivial fixed cost. Table 2 also lacks the same-backbone/same-data model with all three new components disabled, so total method gains cannot be fully separated from the strong backbone and long-horizon training. These omissions matter to the central claim of practical open-ended streaming. W2:The evaluation establishes recurring-content consistency more directly than global style stability on unseen content. Both quantitative long-horizon protocols compare recurrences, and LPIPS/DINO can favor weaker stylization; the strongest Cycle baseline is already close (0.0467 versus 0.0430 LPIPS). Different-video quality scores do not rule out this trade-off on the long streams. The 139k-frame example is qualitative, benchmarks have only five/ten source sequences, and fixed-bank retention is not stress-tested against growing scene diversity. The contribution should be scoped to the demonstrated revisit regimes until style strength and non-recurring-content drift are measured.

Questions for authors

1.Can you report the same-backbone, same-data model with SISM, RSCR and EA-SDM all disabled under the existing protocols? The leave-one-out rows in Table 2 identify marginal effects, but not the total contribution relative to an equally trained base. 2.On existing CAM/Cycle outputs, can you jointly report style adherence and LPIPS/DINO, with source-sequence-level uncertainty, to test whether consistency gains persist at comparable style strength? A style-drift analysis on non-recurring segments of the existing long rollout would also clarify EA-SDM’s role. 3.Please specify calibrator C in Appendix Eq. 13 and the thresholds for memory admission/retrieval. When scene diversity exceeds the M=15 bank and an entry is evicted, how does the system handle a later revisit? Please explain the fallback and provide any existing stress-test evidence. 4.The 96.7 minutes describe source-video duration. What was the actual generation time for the 139,210-frame rollout, with which GPU and resolution? Please report throughput, first-frame/per-chunk latency and peak GPU memory, ideally comparing the same backbone with and without SISM. 5.Can you show representative failures from existing outputs, including unreliable retrieval or incorrect stylization stored in memory? Does changing the style prompt mid-stream require resetting or separating historical memory? If prompt switching is unsupported, please state this limitation. 6.Please give participant and assignment counts, reconciling 420 judgments, 210 distinct case/method-pair combinations and 21 judgments per participant. Did the study assess long-horizon consistency or only stylization quality?

11 comments
AuthorsOpenReview ↗

Author Response (1/n)

Q1: More results of inference cost and ablation study.

A1: Thank you for pointing this out. Following your suggestion, we have added hardware-specific inference measurements and the same-backbone, same-data model with all three components disabled.

  • Inference cost. On a single NVIDIA A100 (80 GB) GPU at 832×480832\times480, EverStyle runs at 7.7 FPS, with a first-frame latency of 1.7 s, a per-chunk latency of 1.56 s, and a peak GPU memory of 38.9 GiB. Regarding the fixed cost you mentioned, all-layer nearest-neighbor retrieval, the additional attention branch, and memory writing take 45, 42, and 30 ms per chunk, respectively. In total, SISM adds about 8% per-chunk time and 1.3 GiB of peak GPU memory over the same backbone without SISM, and this cost stays stable over the 139,210-frame rollout. As 7.7 FPS is below the 24-FPS source rate, the current implementation is not yet real-time (Appx. C); this gap mainly comes from the two-step backbone and VAE rather than from SISM, and closing it is orthogonal to our contributions on long-horizon consistency. A6 reports the full comparison with and without SISM.
  • Same-backbone model with all three components disabled. Relative to this equally trained base, the full method reduces LPIPS by 42% on CAM (from 0.4887 to 0.2830) and by 58% on IVEBench-Cycle (from 0.1027 to 0.0430). A3 provides the training setup and the full results.

We have added these results to Appx. B.9 and Table 2 (highlighted in blue). We hope these additions can help resolve your doubt and make the practical cost and the total contribution of our method clearer.

AuthorsOpenReview ↗

Author Response (2/n)

Q2: Evaluations regarding long-horizon style consistency.

A2: Thank you for this insightful comment. Following your suggestion, we measured the two quantities identified in your comment, style strength and non-recurring-content drift, and we also address the fixed-bank retention and the number of source sequences. In short, EverStyle's consistency is not obtained through weaker stylization, and the requested style remains clearly expressed across different scenes of the 139,210-frame rollout.

  • Style consistency and strength. We supplement LPIPS/DINO with Contrastive Style Descriptors (CSD), a metric specifically designed for image style similarity (Somepalli et al., Measuring Style Similarity in Diffusion Models, ECCV 2024). In a controlled check (Appx. B.3), CSD is clearly higher for pairs that share the target style but differ in content (0.640) than for pairs that share content but differ in style (0.313), so it is suited to comparing style across different content. Comparing the first and final complete cycles of all five methods on the same ten IVEBench-Cycle cases gives:

    MethodMatched-frame CSD ↑\uparrowCross-frame CSD ↑\uparrow
    StreamV2V0.96650.7491
    LiveEdit0.91840.8069
    SANA-Streaming0.71380.6523
    JoyAI-Video-Edit0.89430.7936
    EverStyle0.98020.8736

    EverStyle achieves the highest CSD similarity under both the matched-frame and cross-frame protocols.

    To test whether this consistency is obtained through weaker stylization, we evaluate target-style adherence on the same CAM/Cycle outputs. A blinded GPT-6 Astra evaluator scores each video with a fixed 1–5 rubric at early, middle, and late stages, with identical sampling positions across methods. The results are as follows:

    MethodCAM adherence ↑\uparrowCycle adherence ↑\uparrow
    StreamV2V2.602.53
    LiveEdit2.801.90
    SANA-Streaming4.203.67
    JoyAI-Video-Edit3.673.57
    EverStyle4.204.30

    From these results, we observe that EverStyle matches SANA-Streaming's mean adherence on CAM while achieving better LPIPS/DINO. On IVEBench-Cycle, StreamV2V is the baseline closest to ours in LPIPS (0.0467 versus 0.0430), but its adherence is much lower (2.53 versus 4.30). Thus, our consistency comes with clearly expressed target styles rather than weaker stylization.

  • Quantitative evaluation across different scenes in the 139,210-frame rollout. To make the 139k-frame example quantitative, we evaluate two-second windows near the start, at the midpoint, and near the end of the Singapore video, which depict different scenes; the positions are fixed independently of output quality. We compute the CSD similarity between windows and assess adherence with the same blinded rubric. The results are as follows:

    RolloutEarly–middle CSD ↑\uparrowEarly–late CSD ↑\uparrowMiddle–late CSD ↑\uparrowAdherence: early / middle / late
    Singapore, 139,210 frames0.77920.72400.67345/5/4

    From these results, we observe that the requested style remains clearly expressed in all three windows (adherence 5/5/4), across different scenes and long temporal gaps. A4 further shows that the early–late similarity drops from 0.724 to 0.649 without EA-SDM.

  • Fixed-bank retention under growing scene diversity. The 139,210-frame rollout is our most demanding setting in this respect: it is a 96.7-minute city walk whose scenes change continuously (e.g., a street intersection, a residential street, and a garden waterfront in the three evaluated windows), while the bank is fixed at M=15M=15 and holds only about 0.04% of the latent frames. Under this fixed budget, the requested style remains clearly expressed in all three windows (table above), and the per-chunk cost stays stable. On CAM, the bank likewise holds fewer than 1% of the latent frames of each sequence, and revisits remain more consistent than all baselines (Table 1). A5 details the eviction fallback and the capacity sweep.

  • Number of source sequences. To account for the number of source sequences, we report source-sequence-level bootstrap confidence intervals in A4. On CAM, whose five sequences contain 468 revisit pairs, the improvement over SANA-Streaming, the baseline with matching mean adherence, has a 95% CI that excludes zero (ΔLPIPS=−0.2394\Delta\mathrm{LPIPS}=-0.2394, [−0.3888-0.3888, −0.0958-0.0958]).

We have added the CSD evaluation, target-style adherence results, and long-rollout measurements to Appx. B.3, B.6, and B.7, respectively (highlighted in blue). We hope these additions can address your concerns about style strength and style drift on non-recurring content.

AuthorsOpenReview ↗

Author Response (3/n)

Q3: Report of model with SISM, RSCR, and EA-SDM all disabled.

A3: Thank you for this valuable suggestion. Following your recommendation, we trained the same-backbone, same-data model with SISM, RSCR, and EA-SDM all disabled and evaluated it under the existing protocols. Relative to this equally trained base, the three components together reduce LPIPS by 42% on CAM and by 58% on IVEBench-Cycle. Both models start from the same generator after the 5K-iteration self-forcing DMD stage and use the same long-horizon data, rank-512 LoRA, and 20K training iterations. The base model optimizes only the local DMD objective, with RSCR and EA-SDM disabled and SISM removed in both training and inference. The results are as follows:

ConfigurationTotal ScoreVideo QualityInstruction ComplianceVideo FidelityCAM LPIPSCAM DINOCycle LPIPSCycle DINO
Ours0.758/0.7560.799/0.8090.655/0.6480.821/0.8110.28300.80510.04300.9786
Ours w/o SISM, RSCR, EA-SDM0.744/0.7390.796/0.7970.632/0.6210.803/0.8000.48870.65470.10270.9427

From these results, we observe that, with the backbone, data, and training budget held fixed, adding the three components improves LPIPS/DINO from 0.4887/0.6547 to 0.2830/0.8051 on CAM and from 0.1027/0.9427 to 0.0430/0.9786 on IVEBench-Cycle, and raises the IVEBench-Style Total Score from 0.744/0.739 to 0.758/0.756 (Short/Long). This comparison separates the total contribution of the three components from that of the backbone and long-horizon training, complementing the leave-one-out ablations in Table 2.

We have added this base model to Table 2 and Appx. B.8 (highlighted in blue). We hope this additional comparison can help resolve your doubt and make the total contribution of our method clearer.

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

This paper proposes EverStyle, a long-horizon consistent streaming video stylization framework that addresses historical stylization forgetting and global style drift during extended video generation. Specifically, the authors introduce Source-Indexed Stylization Memory (SISM) to preserve and retrieve historical stylization residuals under a fixed memory budget, Replay-Based Stylization Consistency Regularization (RSCR) to encourage consistent stylization of recurring content under different temporal contexts, and Early-Anchored Stylization Distribution Matching (EA-SDM) to maintain global style consistency throughout long rollouts. Experimental results demonstrate that EverStyle improves long-horizon consistency while maintaining competitive stylization quality.

Critical strengths
  1. The paper addresses a practically meaningful problem with clear motivation. It focuses on long-horizon consistency in streaming video stylization and identifies two distinct but related challenges: historical stylization forgetting and global style drift. This formulation provides a clear motivation for the proposed framework.
  2. The experimental evaluation provides encouraging results. The authors evaluate stylization quality, exact content recurrence, and natural scene revisitation. The paper also includes component ablations, hyperparameter analyses, and a user study, providing empirical evidence for the effectiveness of the proposed approach.
Critical weaknesses
  1. Although the paper emphasizes streaming video stylization and fixed memory usage, SISM introduces additional memory storage, retrieval, and fusion operations across multiple attention layers. It would be helpful to report inference speed (FPS), per-frame latency, GPU memory consumption, and computational overhead under different memory capacities to better demonstrate the practicality of the proposed method for real-time streaming applications. 2.The differences and relationships between RSCR and EA-SDM require further clarification. The paper introduces RSCR and EA-SDM to constrain the reproduction of historical stylizations and global style consistency, respectively. However, the motivation for adopting two different regularization mechanisms and their potential overlap or synergy are not sufficiently discussed. The authors are encouraged to further clarify their differences in optimization objectives and underlying mechanisms, and provide more targeted experiments to demonstrate their complementarity. 3.The evaluation metrics for long-horizon style consistency could be further improved. The paper primarily relies on LPIPS and DINO similarity to evaluate long-horizon consistency. However, these metrics are also influenced by structural and perceptual similarities, making it difficult to isolate style consistency from content preservation. Additional metrics specifically measuring color, texture, or global artistic style drift would provide more direct evidence of the proposed method's effectiveness.
Questions for authors

See Weaknesses

5 comments
AuthorsOpenReview ↗

Author Response (1/n)

Q1: More results of inference cost.

A1: Thank you for your helpful suggestion. Following your recommendation, we measured the inference speed (FPS), per-frame latency, GPU memory consumption, and computational overhead of EverStyle under different memory capacities MM, on a single NVIDIA A100 (80 GB) GPU at 832×480832\times480. In short, at the default capacity (M=15M=15), EverStyle runs at 7.7 FPS, and SISM adds 1.3 GiB of peak GPU memory and about 8% per-chunk time (from 1.44 s to 1.56 s). The per-frame latency is the per-chunk time amortized over the 12 output frames of a chunk. The results are presented below:

MFPSPer-frame latency (ms)Per-chunk time (s)Peak GPU memory (GiB)SISM state (GiB)Retrieval + memory branch (ms/chunk)
0 (SISM off)8.31201.4437.60.000
17.81281.5437.80.1478
77.71291.5538.20.5482
15 (Ours)7.71301.5638.91.0887
317.51321.5940.02.17100
637.31381.6542.34.33130

From these results, we observe that the cost of SISM is modest and bounded. Increasing MM from 15 to 63 lowers the throughput only from 7.7 to 7.3 FPS, while the SISM state grows approximately linearly from 1.08 to 4.33 GiB. At a fixed MM, SISM's storage and per-chunk computation remain bounded as the stream grows, since each attention layer stores at most MM committed frames and one pending candidate. In the 139,210-frame rollout, the average chunk time differs by less than 0.5% between the first and last 10% of the rollout.

Regarding real-time streaming applications, the current throughput (7.7 FPS) does not yet reach real time. This gap mainly comes from the two-step backbone and VAE rather than from SISM (removing SISM only raises the throughput to 8.3 FPS), and accelerating the backbone, which is orthogonal to our contributions on long-horizon consistency, is our future work (Appx. C).

We have added these measurements to Appx. B.9 (highlighted in blue). We hope these additional results can address your concern and provide a clearer understanding of the computational cost and practicality of our method.

AuthorsOpenReview ↗

Author Response (2/n)

Q2: More discussions of the relationship between RSCR and EA-SDM.

A2: Thank you for this valuable suggestion. Following your recommendation, we (1) clarify the differences between RSCR and EA-SDM in their optimization objectives and underlying mechanisms, and (2) add a targeted experiment to demonstrate their complementarity. In short, the two partially overlap, since both use earlier stylized outputs to guide later generation, but RSCR specifies how recurring content should be reproduced, whereas EA-SDM keeps the overall style stable on content that does not recur. Their differences are summarized as follows:

RSCR (Eq. 4)EA-SDM (Eq. 5)
Failure addressedRecurring content is stylized differentlyThe overall style gradually drifts
Applies whenThe same source content is replayedAny content, with no recurrence required
ObjectiveReproduce a specific earlier outputMatch the target distribution through joint DMD evaluation of early and current frames
CorrespondencePosition-wise correspondence between the same source framesNo pixel-wise correspondence required
ConstrainsThe source-to-stylization mapping of replayed contentStylization of current, possibly new content

This is why we adopt two mechanisms: RSCR directly supervises replayed content but provides no explicit target for newly appearing content, whereas EA-SDM extends the constraint across content changes but does not specify the exact earlier output that recurring content should reproduce.

For the targeted experiment, we added the setting with both objectives disabled to complete the four-way comparison (SISM is enabled in all four settings), and we report the early–late CSD style similarity on non-recurring segments of the 139,210-frame rollout. The results are as follows:

RSCREA-SDMTotal Score (Short/Long) ↑\uparrowCAM LPIPS ↓\downarrow / DINO ↑\uparrowCycle LPIPS ↓\downarrow / DINO ↑\uparrowEarly–late CSD on the 139,210-frame rollout ↑\uparrow
OffOff0.748/0.7460.3747/0.74720.0715/0.96310.639
OnOff0.750/0.7500.3416/0.77180.0563/0.97270.649
OffOn0.751/0.7530.3299/0.78150.0628/0.96900.717
OnOn0.758/0.7560.2830/0.80510.0430/0.97860.724

From these results, we observe complementary effects. On IVEBench-Cycle, where identical content recurs, removing RSCR causes a larger LPIPS increase than removing EA-SDM (+0.0198+0.0198 vs. +0.0133+0.0133). On CAM, which includes viewpoint changes, removing EA-SDM causes the larger increase (+0.0586+0.0586 vs. +0.0469+0.0469). On the non-recurring segments, adding EA-SDM alone raises the CSD similarity from 0.639 to 0.717, whereas adding RSCR alone raises it only to 0.649. Using both objectives achieves the best results on all reported metrics, with LPIPS decreasing from 0.3747 to 0.2830 on CAM and from 0.0715 to 0.0430 on IVEBench-Cycle compared with disabling both, which exceeds the improvement from either objective alone. Thus, the two objectives are complementary rather than redundant: although they overlap, each provides benefits beyond those of the other.

We have added this discussion and these results to Appx. B.8 (highlighted in blue). We hope these clarifications and additional experiments can address your concern and make the relationship between RSCR and EA-SDM clearer.

AuthorsOpenReview ↗

Author Response (3/3)

Q3: Evaluations for long-horizon style consistency.

A3: Thank you for this constructive suggestion. Following your recommendation, we added a metric that specifically measures style: Contrastive Style Descriptors (CSD), a learned representation of image style, which covers factors such as colors and textures (Somepalli et al., Measuring Style Similarity in Diffusion Models, ECCV 2024). Under this metric, EverStyle achieves the highest style similarity among all five methods and shows no decline over time.

To verify that CSD isolates style from content, we ran a controlled check (Appx. B.3): CSD is clearly higher for pairs that share the target style but differ in content (0.640) than for pairs that share content but differ in style (0.313).

We then evaluated all five methods on the same ten IVEBench-Cycle cases, comparing the first and final complete cycles under two protocols: matched-frame, which compares all corresponding frame positions, and cross-frame, which compares temporally separated positions to reduce reliance on matching content. The results are as follows:

MethodMatched-frame CSD ↑\uparrowCross-frame CSD ↑\uparrow
StreamV2V0.96650.7491
SANA-Streaming0.71380.6523
LiveEdit0.91840.8069
JoyAI-Video-Edit0.89430.7936
EverStyle (Ours)0.98020.8736

From these results, we observe that EverStyle achieves the highest CSD similarity under both protocols. Its first-to-final cross-frame similarity (0.8736) also remains close to its early-cycle baseline (0.8727, the same comparison within the first cycle), i.e., the established artistic style is maintained over long temporal gaps. To measure style drift over time, we further track the cross-frame CSD at the 25%, 50%, 75%, and final cycles: EverStyle remains the highest at every position, with a net change of only +0.0009+0.0009.

We have added the CSD protocols and results to Appx. B.3 (highlighted in blue). We hope these additional results can address your concern and provide more direct evidence of style consistency.

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

This paper introduces the EverStyle framework, which ensures consistent real-time video stylization across ultra-long video streams. It employs a fixed-capacity memory bank to store and retrieve style variations for previously encountered content (SISM), utilizes replay-based training to help the model reproduce past styles (RSCR), and leverages early generation results as a reference to suppress global style drift (EA-SDM). In tests involving scene revisits and repetitive content, EverStyle demonstrates superior long-range consistency compared to existing methods.

Critical strengths
  1. The problem is clearly decomposed, and the three core mechanisms are well-designed and complementary. SISM addresses the "retrievability" of historical states, while RSCR further ensures their "reproducibility" across different contexts and in the presence of noise. EA-SDM tackles global drift in scenarios where content cannot be reconstructed. Furthermore, ablation studies demonstrate the contribution of each component, enhancing the credibility of the paper.
  2. The experiments are comprehensive, showing significant improvements over existing methods. The approach is evaluated on both the revisiting of natural scenes and the recurrence of content, with results demonstrated on video streams exceeding 139,000 frames. Both qualitative and quantitative results confirm the effectiveness of the proposed method.
Critical weaknesses
  1. Certain mechanisms require more thorough validation. The assumption that style residuals can be transferred across contexts—specifically through nearest-neighbor matching of source features combined with additive residual injection—implicitly posits that historical style variations can be effectively repurposed within the current representation. However, factors such as style, lighting, material, and viewpoint are often entangled within these features; thus, a reliable source match does not necessarily guarantee that the corresponding residual is appropriate. While the EA-SDM mechanism is insightful, it remains unclear whether the constraints imposed by earlier segments on the current output stem primarily from style, or from a video prior that favors content continuity. Furthermore, it is not fully explained how results might be affected by the choice of anchor frames, the method of concatenation, or the specific duration (qq) of the anchor segments.

  2. There is a lack of detail regarding implementation specifics and training costs. For instance, the memory requirements (VRAM usage) of the SISM module and the computational overhead associated with retrieval are not specified. Additionally, metrics such as LPIPS and DINO—which measure general perceptual or feature-level discrepancies between early and late outputs—do not specifically isolate "style drift." In scenarios involving viewpoint changes (e.g., in the CAM dataset), variations in geometry and content may also contribute to these metric scores, raising questions about their reliability for this purpose.

Questions for authors
  1. What are the specific parameter settings—including sampling steps, solver, resolution/output frame rate, chunk/context configurations, weight versions, number of random seeds, and the method for aggregating results—used in this paper and for all baselines?
  2. Further validation of the style residuals and the EA-SDM mechanism.
  3. Should human evaluation be incorporated into the assessment of the videos?
9 comments
AuthorsOpenReview ↗

Author Response (1/n)

Q1: More discussions about SISM and EA-SDM.

A1: Thank you for these insightful comments regarding SISM and EA-SDM. We address the three points below.

  • Transferability of style residuals in SISM under different contexts. We agree that a reliable source match alone does not guarantee an appropriate residual, since style, lighting, material, and viewpoint can be entangled in source features. We clarify that SISM does not reuse a retrieved residual based on source matching alone. First, confidence calibration controls its contribution: beyond matching quality and margin, it considers whether the stored stylization was reliable when written and whether the correction has an appropriate scale relative to the current source (Appx. A, Eq. 13). Second, SISM stores residuals relative to the source rather than historical stylized features (Sec. 3.2, Eq. 3), so only how the content was previously stylized is transferred, while the content, pose, and geometry come from the current source. Third, the calibrated residual is injected through a separate attention branch guided by the current source keys, with the original attention path preserved (Eq. 14). Empirically, removing SISM worsens LPIPS/DINO from 0.0430/0.9786 to 0.0769/0.9587 on IVEBench-Cycle and from 0.2830/0.8051 to 0.3977/0.7237 on CAM, where revisits involve viewpoint changes (Table 2), and A4 further validates the retrieved residuals directly.

  • Origin of the EA-SDM constraint. The early segment contains both style and content, but our design and results indicate that its constraint on the current output mainly acts on style. By design, EA-SDM only adds a detached early stylized segment as a long-range reference (Sec. 3.4): gradients act only on the current frames, which remain conditioned on their own source, so no content correspondence with the early frames is required. Empirically, on windows of the 139,210-frame rollout that depict different scenes, where shared content cannot account for the similarity, EA-SDM raises the early–late CSD style similarity from 0.649 to 0.724 (A4). Moreover, if the constraint came from content continuity, the current frames would be pulled toward the early content and lose fidelity to their own source. Instead, Table 2 shows that neither Video Fidelity nor Instruction Compliance on IVEBench-Style (Short/Long) is reduced by EA-SDM; both are slightly higher (0.821/0.811 vs. 0.807/0.808, and 0.655/0.648 vs. 0.647/0.644).

  • Effects of anchor frames, concatenation, and duration (qq). Following your suggestion, we conducted additional ablations on all three choices: the anchor frames, the method of concatenation, and the duration qq. By default, we sample qq uniformly from {1,…,L−1}\lbrace1,\ldots,L-1\rbrace (L=21L=21 latent frames) and replace the first qq positions of the DMD clip with a qq-frame stylized segment from the earliest portion of the rollout, with gradients acting only on the current frames. We vary one factor at a time under the same training setup, and fix q=10q=10 when varying the placement. The results are as follows:

    FactorSettingTotal Score (Short/Long)CAM LPIPS/DINOCycle LPIPS/DINO
    Default (Ours)earliest segment, placed first, q∼U{1,…,20}q\sim\mathcal{U}\lbrace1,\ldots,20\rbrace0.758/0.7560.283/0.8050.043/0.979
    Anchor framesmost recent qq-frame segment0.750/0.7510.338/0.7730.056/0.973
    Anchor framesrandom qq-frame segment of the generated history0.757/0.7550.304/0.7930.047/0.977
    Durationfixed q=1q=10.754/0.7530.315/0.7870.049/0.975
    Durationfixed q=10q=100.757/0.7550.289/0.8010.044/0.978
    Durationfixed q=20q=200.753/0.7520.308/0.7900.047/0.976
    Concatenationq=10q=10, anchor placed in the middle0.754/0.7520.300/0.7940.046/0.977
    Concatenationq=10q=10, anchor placed last0.752/0.7510.312/0.7870.049/0.976

    From these results, we observe that the default setting gives the best long-horizon consistency.

    1. Anchor frames: The most recent anchor performs close to removing EA-SDM on CAM (0.338/0.773 vs. 0.342/0.772 in Table 2), supporting the benefit of a long-range reference. Random history anchors recover much of the gain, while the fixed early reference performs best.
    2. Duration: A very short anchor (q=1q=1) provides a weak reference, and a very long one (q=20q=20) leaves a single current frame to receive gradients; an intermediate length (q=10q=10) performs close to the default random qq.
    3. Concatenation: At fixed q=10q=10, placing the anchor first outperforms middle or last placement.

Across all variants, the IVEBench-Style Total Score varies by at most 0.008. We have added these results to Appx. B.8 (highlighted in blue). We hope these additional experiments and clarifications can help resolve your doubt and make the roles of SISM and EA-SDM clearer.

AuthorsOpenReview ↗

Author Response (2/n)

Q2: Details regarding implementation and metrics.

A2: Thank you for pointing this out. We address the two points below.

  • More implementation details. Following your suggestion, we measured the memory requirement (VRAM usage) of SISM and the computational overhead of retrieval, and we also report the training cost. In short, SISM adds about 1.3 GiB of peak GPU memory and 0.12 s per chunk (about 8%). The measurements compare the same model with and without SISM on a single NVIDIA A100 (80 GB) GPU at 832×480832\times480. The results are as follows:

    ConfigurationPeak GPU memory (GiB)SISM state (GiB)Retrieval (ms/chunk)Memory branch (ms/chunk)Memory writing (ms/chunk)Total (s/chunk)
    Without SISM37.60.000001.44
    With SISM (M=15M=15)38.91.084542301.56

    From these results, we observe that SISM retains 1.08 GiB of state, with retrieval, the memory branch, and memory writing taking 45, 42, and 30 ms per chunk, respectively. Since each attention layer stores at most MM committed frames and one pending candidate, the memory usage remains bounded as the stream grows. Training on 8 A100 GPUs takes approximately 117 GPU-hours for the 5K-iteration self-forcing DMD stage and 711 GPU-hours for the 20K-iteration long-horizon stage. We have added these measurements to Appx. B.9 and the remaining implementation settings to Appx. A.3 and B.10 (highlighted in blue).

  • Style-focused evaluation. We agree that LPIPS and DINO measure overall appearance consistency and do not isolate style drift. They remain informative for comparing methods because every method is evaluated on the same frame pairs: on IVEBench-Cycle, the paired frames come from identical source frames, so content and geometry are fixed; on CAM, the viewpoint-induced difference between paired frames is shared by all methods, so the gaps between methods reflect how consistently each method stylizes the revisited scene. To further isolate style, following your comment, we additionally report Contrastive Style Descriptors (CSD), a metric specifically designed to measure image style similarity (Somepalli et al., Measuring Style Similarity in Diffusion Models, ECCV 2024). In a controlled check (Appx. B.3), CSD is clearly higher for pairs that share the target style but differ in content (0.640) than for pairs that share content but differ in style (0.313), i.e., it responds to style rather than to content or geometry.

    We then evaluated all five methods on the same ten IVEBench-Cycle cases, comparing the first and final complete cycles under two protocols: matched-frame, which compares all corresponding frame positions, and cross-frame, which compares temporally separated positions to reduce reliance on matching content. The results are as follows:

    MethodMatched-frame CSD ↑\uparrowCross-frame CSD ↑\uparrow
    StreamV2V0.96650.7491
    SANA-Streaming0.71380.6523
    LiveEdit0.91840.8069
    JoyAI-Video-Edit0.89430.7936
    EverStyle (Ours)0.98020.8736

    From these results, we observe that EverStyle achieves the highest CSD similarity under both protocols. Its first-to-final cross-frame similarity (0.8736) is also close to its early-cycle baseline (0.8727, the same comparison within the first cycle), indicating that the established style is preserved over long temporal gaps.

We have added the CSD protocols and results to Appx. B.3 (highlighted in blue). We hope these additional results can help resolve your doubt and make our evaluation of style consistency more complete.

AuthorsOpenReview ↗

Author Response (3/n)

Q3: Specific parameter settings.

A3: Thank you for requesting these details. Following your request, we list the sampling steps, solver, resolution/output frame rate, chunk/context configurations, weight versions, random seeds, and aggregation method for our method and all baselines below, and have added them to Appx. B.10 (highlighted in blue). All methods are evaluated at 832×480832\times480 and keep the source frame rate. We generate one output per method and test case using the seed rules in the table, with no selection across runs.

Method and weightsSamplingChunk and contextSeed
EverStyle · Wan2.2-TI2V-5B, reference-free Wan-Fun variant, generator checkpoint 5,000 + rank-512 LoRA checkpoint 20,0002 distilled flow-matching steps with re-noising. Timestep indices [1000, 100], warped with shift 53 latent frames per block, local attention 15, sink 3, SISM capacity 15Base 0 + one-based case index
StableV2V · release d0f418ac50 steps per stage. Euler ancestral for InstructPix2Pix, DDIM for I2VGen-XL with Ctrl-AdapterFull sequence in one call42 (I2VGen-XL stage)
EchoStyle · Wan2.2-I2V-A14B + high- and low-noise adapters 2b19376640 Flow-UniPC steps per window81-frame windows with 16-frame overlap42 + case ID + zero-based window index
StreamV2V · SD 1.5 + LCM-LoRA-SDv1.5 + TAESD4 LCM steps at indices [30, 35, 40, 45] of a 50-position scheduleFrame-by-frame, one cached frame updated every 4 frames2
LiveEdit · Stage-3 student ar-forcing_002000.pt, release f0269e2a4 self-forcing steps with re-noising, [1000, 750, 500, 250]3 latent frames per block, local attention 9, sink 30
SANA-Streaming · sana_streaming_ar.pth, release 616778bc4 flow-matching Euler steps3 latent frames per regular block, 2 cached blocks, sink enabled, context reset every 20 blocks0
JoyAI-Video-Edit · DiT 0804 for Style and Cycle, 0811 for CAM2 flow-matching Euler steps1 latent frame per chunk, local window of 3 chunks, global sink enabled42
  • Aggregation on IVEBench-Style. We evaluate Short (93 videos) and Long (50 videos) separately. For each split, we average each metric over valid videos, apply the official IVEBench normalization and metric weights, and compute the Total Score as the equally weighted average of Video Quality, Instruction Compliance, and Video Fidelity. StableV2V completes 27 of the 50 Long videos, so its Long scores are averaged over those 27 videos (†\dagger in Table 1).
  • Aggregation on the long-horizon benchmarks. On IVEBench-Cycle, we compare the first cycle with later cycles at 25%, 50%, 75%, and the final position, matching frames by their within-cycle positions; LPIPS and DINO are averaged over matched frames, then equally over the four positions and the ten cases. On CAM, each of the 117 annotated anchors across five 7,601-frame videos has four later revisits, and we average LPIPS and DINO equally over all 468 frame pairs.

We hope these details can help resolve your doubt and make our evaluation protocol clearer.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Summary (for Reviewers and AC)

We thank all four reviewers and the AC for their time and constructive comments. As the Author-Reviewer discussion period closes, we provide a brief summary of our responses and the revised paper (changes highlighted in blue).

In our responses, we addressed each weakness and question raised by the four reviewers with additional analyses and experiments on style consistency, inference cost, limitations, and failure cases, as well as a high-level overview of the method (Appx. A.1, B.3, B.6, B.7, and B.11). Reviewer i22S appreciated "the additional experiments and careful revision", and following their further suggestion, Fig. 16 now also includes more cases with different views and additional intermediate frames.

The results further support the strengths highlighted across the four reviews: EverStyle addresses a practically meaningful and clearly decomposed problem through three complementary components, and substantially improves long-horizon consistency while maintaining competitive stylization quality.

We hope this brief summary helps the reviewers and the AC more easily connect the initial reviews with the newly provided evidence. We sincerely appreciate your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.