acceptodds
Under review as a conference paper at ICLR 2027

Cache Homeostasis: Diagnosing and Controlling Feature Drift in Autoregressive Video Diffusion

Abstract

Autoregressive video diffusion supports streaming generation by producing video segments sequentially and reusing historical features through a key–value cache (KV Cache). However, cached feature statistics can drift during long rollouts, potentially affecting video quality. To understand this problem, we study how cached value features evolve during long video generation. We find that their energy becomes increasingly concentrated in a small subset of feature directions. Crucially, this change is mainly driven by the growth of a component shared across tokens, rather than by the disappearance of differences between them. To distinguish these effects, we derive an exact decomposition that separates the contributions of shared energy, alignment, and centered variation. Experiments show that the shared component explains most of the observed concentration increase and that its growth is associated with lower aesthetic quality. Guided by this finding, we propose Cache Homeostasis, a training-free method that regulates newly written cache features relative to an early reference. Its lightweight version controls the shared component, while its full version also regulates excess energy in selected centered directions. Experiments demonstrate that reference-directed intervention suppresses abnormal cache growth, while the full version improves text consistency over the lightweight version, which introduces little computational cost.

Primary area:
generative models
Keywords:
Autoregressive Video Diffusion, KV Cache, Long Video Generation, Feature Drift, Spectral Concentration, Training-Free Inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_1JwnOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the feature drift in the key-value (KV) cache during long-horizon autoregressive video generation. By deriving an exact mathematical decomposition, the authors reveal that the observed spectral concentration of cache features is predominantly driven by the amplification of a shared component across tokens, rather than the loss of token-specific variation. Guided by this diagnosis, they propose "Cache Homeostasis," a training-free inference-time intervention that regulates newly written cache values against an early reference state. Extensive experiments demonstrate that the lightweight Mean controller effectively suppresses abnormal cache growth with minimal computational overhead, while the Full controller offers further text-consistency gains at the cost of higher computation and subject consistency.

Critical strengths
  • The paper honestly reports negative results, trade-offs, and unvalidated hypotheses, while the appendix provides exhaustive raw data, algebraic consistency checks, and reproducibility guarantees, setting a remarkably high standard for empirical rigor.
Critical weaknesses
  • The presentation quality is severely compromised by rudimentary figures and overwhelmingly dense, difficult-to-read tables, which strongly appear to have been directly generated or hastily formatted by AI tools without adequate human refinement—a suspicion directly corroborated by the authors' own admission in the "AI USE STATEMENT" regarding their reliance on AI for figure and table preparation.

  • Lack of Causal Link and Limited Predictive Power: Although the paper establishes a raw correlation between shared-energy growth and aesthetic degradation, the authors themselves acknowledge that this association significantly attenuates after temporal detrending.

Questions for authors

See weakness

2 comments
AuthorsOpenReview ↗

Clarification Regarding AI Usage

Dear Reviewer,

We sincerely thank you for your careful assessment of our manuscript and for your valuable feedback. We would like to respectfully clarify that, as explicitly stated in the AI Use Statement on page 11 (Lines 540–545) of our manuscript, generative AI tools were used to assist with manuscript translation and language polishing. All AI-assisted text was carefully reviewed by the authors to verify its accuracy, ensure consistency in technical terminology, and preserve the intended scientific meaning.

We fully recognize the importance of academic integrity, transparency, and scientific rigor. The authors take full responsibility for the scientific content, mathematical derivations, experimental results, and conclusions presented in the manuscript.

We sincerely appreciate your time and consideration, and thank you again for your valuable feedback.

Best regards, The Authors

Thank you for recognizing the rigor of our empirical reporting and reproducibility materials. We respectfully ask that the soundness assessment distinguish the validity of our diagnostic and control results from the strength of causal claims about perceptual quality. Our two central contributions can be evaluated independently of a causal interpretation of the energy-quality correlation.

1. An exact diagnostic distinction with quantitative validation. Increasing uncentered feature concentration does not uniquely imply that token-specific variation has collapsed. Our decomposition separates shared energy, mean-covariance alignment and weighted centered variation, allowing their contributions to an observed concentration change to be identified exactly. This resolves an ambiguity that inspecting an uncentered spectrum alone cannot resolve.

The fixed 12-prompt, two-seed study provides quantitative evidence: the shared term accounts for 97.05% of the prompt-averaged concentration increase and more than half of the increase in 23/24 trajectories, with the counterexample retained. The contribution is therefore both a checkable mathematical diagnostic and an empirical characterization of the studied rollouts.

2. A minimum-change control result with an explicit scope. At alpha=1, Mean is the unique update minimizing the Frobenius perturbation among all value matrices whose token mean lies within the specified reference ball (App. B.2). It preserves the centered component, and hence every within-block pairwise token difference, in exact arithmetic. This connects the diagnosed component to a precisely defined intervention while avoiding unnecessary changes to centered variation.

The optimality statement concerns a fixed constraint at a single cache write; perceptual consequences require separate empirical evaluation. Nevertheless, this local optimization result stands independently of whether the observed energy-quality association is causal. Together, the decomposition and projection provide a mathematical basis for distinguishing what changes and controlling that component with minimal perturbation.

3. Interpreting the temporal evidence precisely. The attenuation after detrending limits the causal and predictive interpretation of the energy-quality association. We report it explicitly for that reason. It does not invalidate the decomposition, the measured attribution or the minimum-change property. Our intervention experiments separately evaluate regulation of cache statistics and endpoint-specific output effects, including improvements and unfavorable tradeoffs. We do not present the correlation as proof that shared growth causes perceptual degradation, or as a validated predictor of intervention benefit. We respectfully ask that these distinct claims be assessed on their respective evidence.

4. AI disclosure and presentation. Our account follows the complete AI Use Statement in the manuscript (p. 11). Quantitative results are computed from recorded model outputs through executable analysis; AI-assisted visual descriptions are not human evaluations. The authors are responsible for the final content, claims and artifacts. The use or visual appearance of AI-assisted materials does not, by itself, establish an error in the derivations or results. We welcome specific technical concerns that can be checked against the proofs, recorded measurements and analysis code.

We will also improve readability by focusing the main narrative on the 50-prompt benchmark and fixed 12-prompt attribution study, moving developmental probes and legacy variants to the appendix, simplifying the main tables, and improving figure labels, typography and captions.

5. Ethics clarification and rating reconsideration. The ethics flag currently has no accompanying explanation. Could you please identify the concrete ethical concern, together with the relevant passage, practice or supporting evidence? If the concern relates to AI assistance, specifying the practice at issue would allow us to respond substantively and accurately.

We respectfully ask you to reconsider the rating and soundness assessment in light of the independently checkable diagnostic decomposition, the minimum-change control result, their empirical evaluation, and the clarified scope of our claims. We appreciate your consideration and welcome specific remaining concerns.

Reviewer_RD6gOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper attributes increasing value-cache spectral concentration to amplification of the shared token mean and proposes two training-free controllers: Mean constrains mean drift against an early reference; Full additionally suppresses selected centered-covariance excess. Results show quality gains with meaningful tradeoffs.

Critical strengths
  • The exact decomposition distinguishes shared amplification from loss of token-specific variation; Mean has a sound minimum-change interpretation.
  • 500 trajectories, matched conditions, prompt-level pairing, and multiple-testing correction support credible comparisons form careful main evaluation.
Critical weaknesses
  • The headline concentration attribution rests on four audits. Energy–quality correlations weaken substantially after temporal adjustment, leaving unclear when the diagnosis predicts intervention benefit.
  • Mean/std alignment is tested on only six development prompts; closely related TetherCache lacks a matched comparison.
Questions for authors
  • Can blinded continuous-video evaluation verify prompted motion and scene-transition correctness, especially when the early reference becomes outdated?
  • The authors should consider discussing/citing some other memory compression in video generation papers, including:
    • Long-term Consistent World Simulation with Memory. (NeurIPS 2025)
    • Mixture of Contexts for Long Video Generation. (ICLR 2026)
    • SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer. (ICLR 2026)
    • Pretraining Frame Preservation in Autoregressive Video Memory Compression. (arXiv 2026)
    • Moga: Mixture-of-groups attention for end-to-end long video generation. (ICLR 2026)
    • StreamMem: Query-Agnostic KV Cache Memory for Streaming Video Understanding. (arXiv 2025)
2 comments
AuthorsOpenReview ↗

Expanded Attribution and Alignment Evidence, Motion Evaluation Clarifications, and Related Work

We thank the reviewer for helping us strengthen coverage and clarify the evidence. For a fixed mean-reference ball at alpha=1, Mean uniquely minimizes Frobenius change, preserving centered differences in exact arithmetic; Full adds an optional centered correction (App. B.2).

1. Attribution, limited decline, and prediction The Native (uncorrected generation) attribution study uses 12 prompts and two matched realizations per prompt: 24 trajectories and 67,680 measurements. The prompt is the independent unit; layers and repeated measurements are not replicates. Cohort shares are ratios of averaged component changes, with equal prompt weights. Shared, alignment, and weighted centered terms account for 97.05%, 3.69%, and -0.74% of cohort concentration growth. Shared dominance holds in 23/24 trajectories; the retained counterexample has a 43.82% shared contribution. This finite cohort does not establish prevalence across all 100 original trajectories (App. C.12, Table 14).

No fixed-cohort prompt meets the unchanged aesthetic-change rule, > -0.05. Post-hoc screening of the original data finds 2/50 qualifying prompt means and 3/100 trajectories. We retain both realizations per prompt: four trajectories and 11,280 measurements, including the fourth realization below threshold (-0.05039). Dancers improve aesthetically while shared growth contributes 92.05%; skateboarder shared/alignment/weighted-centered shares are 7.79%/52.86%/39.36%, with opposite-signed concentration changes. Its 83.09% shared share describes a decrease. These outcome-selected, Native-only cases remain separate from the fixed cohort, are not pooled, and do not predict Mean/Full benefit (App. C.13, Table 15).

The same-trajectory energy-quality correlation weakens from -0.697 to -0.314 after detrending and -0.140 after differencing, with comparable reverse-time results. These associations establish neither causality nor a validated benefit predictor.

2. Mean/std alignment and Full's role The follow-up retains 24 new alignment videos and 48 historical Mean/Full controls, with paired inference over 12 prompts. All six primary Full-minus-alignment endpoints fail Holm correction (minimum adjusted p = 0.1843). A positive marginal text interval does not establish corrected significance. Full is optional; these results establish neither superiority nor equivalence (App. E.2, Table 38).

Prompts were fixed before new outputs but after original outcomes were known; this is not an independent holdout or a completed 50-prompt comparison. New A800 runs reuse historical A100 controls, limiting hardware-identical operator-only inference. The original six-prompt results remain development evidence.

Our alignment baseline adjusts newly written values at the same write site. TetherCache/TAME aligns recalled keys/values using a trusted sink-and-memory pool. The baseline does not reproduce its retrieval, reference construction, or full correction; a matched full-system comparison is not complete (App. E.1).

Original Full-versus-Mean results show text/subject-consistency tradeoffs (Table 16). Separate timing over two prompts and eight completed exclusive-A100 runs measures generation-only overheads of +1.7% for Mean and +56.6% for Full, excluding loading, text encoding, decoding, and diagnostics. These costs are not inferred from cross-device alignment runs (Table 39).

Sink+Mean improves background/smoothness but fails its aesthetic-or-text criterion; imaging has a negative estimate (Table 3).

3. Continuous motion and aging references We have not conducted the proposed blinded continuous-video human study. Existing evaluation covers all 500 videos, all 3,837 continuously decoded frames per video, and all 20 windows across subject-motion, camera-motion, scene-change, and stable-scene strata (Apps. D.1-D.2). Automatic consistency, smoothness, and dynamic degree cannot establish intended action/transition completion; dynamic degree measures activity, not motion fidelity, and lower activity is not a quality gain.

A fixed early reference may suppress intended transitions; the local minimum-change guarantee does not certify its later semantic suitability (Secs. 5.1/7). Fixed examples in App. D.6, Figs. 6-7 (pp. 36-37), show early/middle/late views but cannot replace continuous assessment.

4. All six related works Sec. 2 (p. 3), with references on pp. 12-13, now discusses WorldMem; Mixture of Contexts (MoC); SANA-Video; TinyHistory (the current title of Pretraining Frame Preservation in Autoregressive Video Memory Compression, arXiv:2512.23851v6); MoGA; and StreamMem. They address memory retrieval/routing, compact history, sparse attention, and streaming understanding. Our focus is newly written-value statistics at fixed weights/cache capacity. This is conceptual positioning: compositions with these architectures have not been evaluated.

I do not have many problem with this paper and intend to keep my score.

Reviewer_PHn7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how stored attention features change during long video generation. It finds that a component shared by all tokens grows and makes the features appear less diverse, even when differences between tokens remain. The authors provide a mathematical way to separate these effects. They then propose two training-free methods: Mean keeps the shared component close to an early reference, while Full also limits growth in selected directions of the remaining features.

The main evaluation uses 500 four-minute videos across 50 prompts, two seeds, and five methods. The methods control feature growth and improve some quality scores, but introduce tradeoffs.

Critical strengths

The paper provides a useful explanation of what changes inside the cache. Separating shared growth from changes in token differences makes the diagnosis more precise. This is the clearest contribution and could improve how future work studies feature drift.

The main evaluation is substantial and carefully reported. It includes all 500 assigned videos, uses prompts as the independent samples, and reports unfavorable results as well as improvements. This makes the reported tradeoffs more credible.

Critical weaknesses

The practical benefit over the simple Sink baseline is unclear. Sink achieves better aesthetic, image-quality, and text-consistency scores, although the proposed methods retain higher subject and background consistency. Full also adds about 57% to generation time. Without stronger evidence about which tradeoff users prefer, the case for adopting the proposed methods is limited.

The connection between the diagnosed feature growth and poor video quality remains weak. The association becomes much smaller after accounting for time trends, and tighter control of feature growth does not consistently produce better quality. The additional-generator tests also do not establish reliable benefits. This limits the broader significance of the proposed diagnosis and correction.

Questions for authors

Do the existing main cohort logs allow you to check whether the shared component explains most of the concentration increase across all Native runs, including those with little quality decline? This would clarify how broadly the headline diagnosis applies.

Do you have results comparing Mean and Full with simple mean and standard deviation alignment beyond the six development prompts? Such results could change my assessment of whether the proposed correction provides a meaningful advantage.

2 comments
AuthorsOpenReview ↗

Additional Attribution and Alignment Results, Quality Tradeoffs, and Computational Cost

Thank you for these questions. Native is the uncorrected generator; Mean corrects shared token means, Full also corrects centered covariance, and Sink retains early cache frames.

1. Attribution and its limits. Equations (5)-(6) decompose concentration changes into shared, alignment, and weighted centered terms, without identifying causality or inevitable quality loss. At alpha=1, Mean uniquely minimizes Frobenius change for a fixed early-reference ball, preserving centered differences in exact arithmetic (App. B.2). This does not guarantee the reference's later semantic suitability or video quality. Full remains optional.

The fixed Native study covers 12 prompts, two matched realizations per prompt, 24 trajectories and 67,680 measurements. We average measurements within trajectories, realizations within prompts, and weight prompts equally. Shared/alignment/weighted-centered shares are 97.05%/3.69%/-0.74%. Shared accounts for over half the increase in 23/24 trajectories; the exception is 43.82%. This is not prevalence across all 100 original Native trajectories (App. C.12, Table 14, pp. 24-25).

For aesthetic change > -0.05, only 2/50 original prompt means and 3/100 trajectories qualify. We separately inspect both matched realizations of each qualifying prompt: four trajectories and 11,280 measurements, selected post hoc by outcome. Dancers show an aesthetic gain with prompt-mean shared share 92.05%; skateboarder shares are 7.79%/52.86%/39.36%, with opposite-sign concentration changes. Its individual 83.09% shared share describes a concentration decrease. Shares use signed component/total changes; prompt shares follow component averaging. These cases are analyzed separately, not pooled with the 12-prompt cohort, and do not test controller benefit (App. C.13, Table 15, pp. 26-27).

2. Simpler mean/std alignment. We add 24 mean/std-aligned videos on the 12 prompts, compared with 48 historical Mean/Full controls. Two matched realizations per prompt are averaged, with the prompt as paired unit (n=12). This matches channelwise early-reference means/stds at the same write position; it does not reproduce full TetherCache/TAME or isolate centered direction at matched correction magnitude. All six Full-minus-alignment quality tests fail Holm correction (minimum adjusted p=0.1843). Nonsignificance establishes neither superiority nor equivalence; Full remains optional. Selection preceded new outputs but followed known original outcomes: this is no independent holdout, and new A800 runs use historical A100 controls (App. E.2, Table 38, pp. 50-51).

3. Practical value, composition, and cost. The original evaluation uses 50 prompts and two matched realizations per prompt (paired n=50). Mean-minus-Native aesthetic/text differences are +0.0238/+0.0401, with unadjusted 95% intervals excluding zero, but background consistency falls 0.0080; subject/imaging means fall with intervals crossing zero. Against Sink, Mean has higher subject/background consistency and smoothness means, but lower aesthetic/imaging/text means. These post hoc contrasts do not establish general preference for Mean (Table 19, p. 30).

Sink+Mean adds 100 treatment videos and eight fresh Sink controls (108 new videos); 92 Sink controls are reused. All 50 reused prompts retain both realizations (paired n=50). Only background/smoothness pass the six-test Holm family (p=0.0016/0.0121); subject p=0.0521 does not. Imaging is -0.02496, with interval [-0.05144, 0.00152], so meaningful harm is not excluded. The aesthetic estimate is slightly negative; neither aesthetic nor text passes correction, and the prespecified usefulness criterion fails (Table 3, p. 9).

Cost is measured separately: two development prompts, one matched realization per prompt, eight completed A100 runs across four methods. Mean/Full generation-only overheads are 1.7%/56.6%; loading, encoding, decoding and diagnostics are excluded. This is not end-to-end throughput, and Sink+Mean incremental cost is unmeasured (Table 39, p. 52). Original Full-minus-Mean results also trade a corrected text gain for lower subject consistency (Table 16, p. 28).

4. Quality association and transfer. Native log-common-energy/aesthetic correlation weakens from -0.697 to -0.314 after detrending and -0.140 after differencing; similar reverse-time results indicate shared temporal trend. Between-prompt concentration-change/aesthetic-change correlations are positive, not evidence that greater concentration growth predicts greater aesthetic loss. Neither analysis establishes causation or predicts intervention benefit (Sec. 4.2). Transfer is mixed: no corrected quality gain in Causal Forcing; MAGI's small development scans show scene-dependent tradeoffs, including lower aesthetic/text under Mean on the scene-change prompt. MAGI provides no independent significance test (Sec. 6.5).

AuthorsOpenReview ↗

Clarification of the Contributions and Practical Significance

We sincerely thank the reviewer for the thoughtful assessment and constructive feedback, particularly for recognizing the value of distinguishing shared feature growth from changes in token-level differences. Building on our previous response and additional experiments, we would like to further highlight the scientific contributions and practical significance of our work.

First, our mechanistic analysis provides a quantitative understanding of feature concentration in long-horizon autoregressive video diffusion. Through an exact mathematical decomposition, we distinguish the contributions of shared, alignment, and centered components to feature concentration dynamics. Our additional analysis shows that the shared component accounts for 97.05% of the prompt-averaged concentration change in the examined Native trajectories and explains more than half of the increase in 23 out of 24 trajectories. These findings provide a more detailed characterization of internal feature evolution than aggregate concentration measurements alone, offering a useful analytical framework for investigating cache dynamics in long-horizon generation.

Second, Cache Homeostasis translates these analytical findings into a mathematically motivated intervention. Under the specified fixed-reference constraint, Mean correction at α = 1 uniquely minimizes the Frobenius perturbation while preserving centered token differences in exact arithmetic. This property provides a principled basis for selectively regulating shared feature statistics. In our paired evaluation, Mean correction improves aesthetic and text-alignment scores over Native by +0.0238 and +0.0401, respectively, with unadjusted 95% confidence intervals excluding zero. Furthermore, the measured generation-only overhead of Mean correction is 1.7% in our A100 benchmark, demonstrating its computational efficiency in the evaluated setting.

Third, the broader contribution lies in connecting mechanistic analysis with targeted intervention. Rather than treating feature concentration solely as an empirical observation, our framework provides an interpretable decomposition, identifies a prominent component of the observed dynamics, and develops a corresponding correction mechanism with well-defined mathematical properties. Together, these elements establish a coherent approach to analyzing and regulating cache feature evolution in autoregressive video diffusion.

We sincerely appreciate the reviewer’s valuable feedback and careful consideration. We hope that these clarifications, together with the additional experimental evidence provided in our previous response, help convey the scientific contributions and practical relevance of our work.

Reviewer_J2tvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies KV-cache drift during 240-second rollouts in autoregressive video diffusion. Comparing the uncentered second-moment spectrum with the centered covariance spectrum of newly written value vectors (at fixed token counts), the authors derive an algebraic identity that splits spectral concentration into three additive terms: shared token-mean energy, mean-covariance alignment, and weighted centered variation. Audits on Self-Forcing (Wan2.1-T2V-1.3B) show that the rise in concentration comes mostly from growing token-mean energy ("shared-component amplification") rather than collapsing token-to-token differences. Based on this observation, the authors introduce Cache Homeostasis, a training-free rule that edits newly written values using early chunks (chunks 2–6) as a reference. The Mean variant projects the token mean onto a ball around the reference mean; Full also shrinks variance along the top-rr eigenvectors of Σ−Gref\Sigma - G_{\text{ref}}. Experiments cover a 50-prompt (500-video) 240s benchmark on Self-Forcing, combinations with attention sinks (Sink+Mean, Sink+Full), per-channel mean/std baselines, wall-clock timing, and transfer runs on Causal Forcing and MAGI-1 4.5B.

Critical strengths
  1. Clear diagnostic split between mean shift and covariance collapse: Separating uncentered second moments from centered covariance is a neat, simple way to inspect KV caches. Eq. (6) shows that over 91% of the concentration increase in sink-free Self-Forcing comes from token-mean growth rather than a loss of pairwise token diversity. Holding token counts fixed to avoid participation-ratio bias, tracking RNG states, applying Holm corrections, and reporting negative results openly make the empirical numbers trustworthy.

  2. Simple, cheap Mean update on sink-free rollouts: On uncontrolled sliding-window Self-Forcing (Native), projecting the token mean back to the reference ball removes sustained shared-energy spikes (0/100 vs. 86/100 trajectories) and lifts aesthetic (+0.024) and text consistency (+0.040) scores while adding only 1.7% generation time.

Critical weaknesses
  1. Outperformed and rendered redundant by standard attention sinks: Keeping a few initial sink frames is already standard, zero-cost practice in streaming video diffusion (including the tested Causal Forcing model). Once a 4-frame sink (Sink) is turned on in Self-Forcing, three facts leave little room for Cache Homeostasis: (1) sustained shared-energy growth mostly disappears on its own (dropping from 86/100 in Native to 2/100 in Sink, with pre-write energy growth at 1.192 vs. 1.258 under Mean); (2) Sink beats both Mean and Full by a wide margin on 240s video quality (Aesthetic: 0.553 vs. 0.436/0.444; Imaging: 0.688 vs. 0.614/0.586; Text: 0.238 vs. 0.162/0.186), whereas Mean and Full still suffer from heavy late-stage striping and distortion (Figure 6); and (3) stacking the method onto sinks does not help—Sink+Full hurts aesthetic (-0.040) and imaging (-0.083) quality, and Sink+Mean fails the authors' own usefulness criterion with no significant gain in aesthetic or text scores. In short, the method neither beats nor complements basic attention sinks.

  2. Flawed static-reference assumption, unconvincing Full variant, and poor generalization: Tying value statistics to chunks 2–6 assumes early feature distributions should hold forever, which clashes with natural camera pans and scene cuts in multi-minute videos (contributing to my Soundness score of 2, alongside the fragmented multi-cohort setup). Meanwhile, Full slows generation by 56.6% due to per-layer, per-chunk eigendecompositions, yet trades better text alignment (+0.024) for worse subject consistency (-0.011) over Mean and fails to significantly beat simple per-channel mean/std matching (minimum Holm p=0.1843p = 0.1843). Finally, neither the post-startup shared-energy surge nor any quality gain carries over to Causal Forcing (all Holm p=1.0p = 1.0) or MAGI-1 4.5B. Shared-component amplification looks more like a side effect of running Self-Forcing without sinks than a general bottleneck in autoregressive video diffusion.

Questions for authors
  1. Effect of removing native sinks in Causal Forcing: Causal Forcing shares the Wan2.1-1.3B backbone with Self-Forcing, yet its common-energy fraction stays flat (~0.203) over 240s and neither controller helps. Because the tested Causal Forcing pipeline keeps a 3-frame sink by default (Line 3176), what happens if you turn off those 3 sink frames? Does shared-component amplification come back, or does Causal Forcing's training prevent it even without sinks?

  2. Why sinks control drift better than value projection: Sink keeps pre-write shared energy lower than Mean (1.192 vs. 1.258) and produces much cleaner 240s videos (0.553 vs. 0.436 Aesthetic). Why does attending to four untouched early KV frames stabilize generation so much better than clamping newly written value means?

  3. Cleaning up the cohort structure (presentation suggestion): Reading the main text requires tracking seven different cohorts and legacy variants (e.g., Ref. mean in Figure 1c vs. Mean in Figure 1d). Could the main text focus purely on the 50-prompt benchmark and the 12-prompt attribution runs, moving earlier developmental probes to the appendix?

2 comments

Thank you for your detailed review and for recognizing the diagnostic distinction and the rigor of our evaluation. Your questions help us clarify the relationship between sink retention and value correction, the scope of our evidence, and the presentation of the paper.

1. Causal Forcing without its native sinks (Q1). We do not have time to do a completed, validated sink-off comparison to report. Both the Causal Forcing quality comparison and the additional Native diagnostic study retain its native three-frame sink. Consequently, these experiments cannot distinguish the contribution of training from that of sink retention, and they do not establish that training alone prevents shared-component amplification. We will make this limitation explicit and describe the result as applying to the tested native pipeline.

2. Why Sink stabilizes generation more strongly than Mean (Q2). Sink and Mean affect different aspects of generation. Within the same 21-frame cache budget, Sink retains four early frames and 17 recent frames, providing continued access to early keys and values. Mean retains the recent-frame window and modifies only newly written values; it does not restore early content or directly change the keys.

Retaining early context can influence query-dependent attention throughout generation, whereas Mean constrains one component of newly written features. At fixed queries and keys, shifting a value block’s mean changes its attention contribution in proportion to the attention mass assigned to that block. This local correction therefore does not reproduce the broader effects of retaining early context.

This distinction offers a plausible explanation for Sink’s stronger stabilization, although our experiments do not isolate the causal mechanism behind the performance gap. The pre-write growth ratios you highlight—1.192 for Sink versus 1.258 for Mean—also show that tighter post-write regulation does not necessarily produce better subsequent features or higher perceptual quality.

3. The fixed early reference and the role of Full. Mean constrains the token mean to a ball determined by early-reference statistics, while preserving within-block token differences in exact arithmetic. At alpha=1, it is the minimum-Frobenius-change update satisfying that constraint (App. B.2). It does not require the entire early feature distribution to remain unchanged.

Nevertheless, a legitimate camera or scene transition can move the mean outside the reference ball. The minimum-change property does not establish that the reference remains semantically appropriate throughout a rollout. We will explain this limitation alongside the controller and distinguish it from the diagnostic decomposition, which remains exact for changing means and covariances.

Full is an optional extension for studying centered correction. Its 56.6% generation overhead, text/subject-consistency tradeoff, and lack of corrected superiority over mean/std alignment limit the case for using it as the default. Mean is the primary lightweight controller, with measured generation overhead of 1.7%.

4. Practical scope and composition with Sink. Sink achieves higher aesthetic, imaging and text scores than Mean in this benchmark. Mean has higher subject/background consistency and smoothness scores, but these endpoint-specific differences do not establish an overall preference for Mean.

Adding Mean to Sink improves background consistency by 0.00397 and smoothness by 0.00202 after Holm correction (p=0.0016 and 0.0121; Table 3). However, neither aesthetic nor text consistency satisfies the prespecified usefulness criterion, and the negative imaging estimate leaves possible harm unresolved. We will describe these findings as selective automatic-metric gains, without claiming broad perceptual improvement from composition.

The contribution supported by the evidence is a component-level diagnosis and a corresponding minimum-change control design. In the fixed 12-prompt study, the common term explains 97.05% of the averaged concentration increase and more than half in 23/24 trajectories. These findings characterize the studied regime; they do not establish shared amplification as a universal bottleneck or identify when intervention will improve quality.

We appreciate your careful questions and the opportunity to clarify the method’s scope and limitations.

Thank you again for your careful consideration. Building on our previous response, we would like to offer a brief clarification.

The paper provides two complementary, independently checkable contributions. First, the exact decomposition distinguishes shared-component amplification from changes in centered variation, resolving an ambiguity that an uncentered spectrum alone cannot resolve. The fixed 12-prompt study supports this distinction quantitatively: the shared term explains 97.05% of the prompt-averaged concentration increase and more than half in 23/24 trajectories. Second, at alpha=1, Mean provides the unique minimum-Frobenius-change update under the stated reference-ball constraint, preserving within-block token differences in exact arithmetic. Together, these results connect a precise diagnosis to a targeted correction.

Mean also offers specific practical advantages: it retains the full 21-frame recent-history window without reserving early KV frames and adds only 1.7% generation overhead. In the main benchmark, it achieves higher subject/background consistency and smoothness means than Sink. When added to Sink, it further improves background consistency and smoothness after Holm correction (p=0.0016 and 0.0121). These are endpoint-specific benefits that complement the diagnostic and mathematical contributions.

We hope this clarification helps convey the value of the work alongside the quality comparisons discussed in our response.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.