acceptodds
Under review as a conference paper at ICLR 2027

Delta Forcing: Trust Region Steering for Interactive Autoregressive Video Generation

Abstract

Interactive streaming video generation requires a causal model to react to evolving event conditions while preserving the visual trajectory it has already established. A fundamental challenge lies in balancing reactivity and stability: models must respond promptly to new events while maintaining temporal coherence over long horizons. Existing approaches typically distill high-quality bidirectional diffusion models into autoregressive generators and further adapt them via streaming long-tuning; however, they often suffer from identity, layout, and scene drift when event switches. We identify the cause of this failure as conditional bias, where the teacher can produce denoising guidance that is semantically valid for the new event but incompatible with the student’s realized history, thereby causing distribution matching to propagate prompt-consistent yet trajectory-inconsistent modes. We propose Delta Forcing, a transition-calibrated steering objective for a distribution-matching distillation. For each generated chunk, Delta Forcing compares the feature-space delta of the teacher’s denoised estimate with the delta of the generator’s history-grounded rollout, retaining full teacherdriven distribution matching when the transitions agree and shifting toward trajectory-continuity regularization when they diverge. This adaptive supervision restores the balance between reactivity and stability, enabling coherent long-term generation under dynamic conditions. Extensive experiments demonstrate that Delta Forcing improves long-horizon stability while maintaining responsiveness to evolving events.

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

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_h1izOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates temporal drift in autoregressive video generation under changing event prompts. The proposed method adjusts DMD supervision according to the feature displacement of consecutive teacher predictions. When this displacement becomes unusually large, a student momentum loss receives greater weight to preserve the recent trajectory. The experimental results show improved subject and background consistency while maintaining competitive prompt alignment.

Critical strengths

The proposed momentum loss encourages consistency between consecutive feature displacements, allowing temporal continuity to be maintained without directly forcing adjacent frames to remain similar. This design is appropriate for interactive generation, where the scene should remain consistent while the requested action changes.

Critical weaknesses
  1. The attribution of drift to teacher conditional bias rests on one qualitative example (Fig. 1 right) and the PCA plots in App. A. Sec. 3 argues that early onset rules out error accumulation, but that does not show the jump originates in the teacher. The reliability probe in Eq. 4 is built from the student's own chunk (Alg. 1, line 19d), so the displacement in Eq. 5 already contains the student's displacement, and the gate cannot tell a teacher mode change from a student jump. Lower drift after gating is consistent with either explanation.
  2. The sensitivity sweep in Fig. 4 (right) shows that gating more often or weighting the fallback more lowers Dynamic Degree, and App. B concedes that motion starting at the prompt boundary is temporarily downweighted. Long-CLIP in Table 1 is averaged over ten-second windows, so a short delay in realizing the new prompt would be invisible, and no reported metric measures action onset.
Questions for authors
  1. The Design 1 and Design 2 removal experiments in Fig. 4 (left) provide qualitative comparisons, but their effects are not quantified using the VBench metrics. Please report Table 2 metrics for those variants and for a constant gate weight, and give the values of the EMA rate, warm-up length, clip factor, and probe level in Alg. 1, which are not stated anywhere.
  2. Please report the action-onset delay relative to LongLive and MemFlow, for example the first chunk at which Long-CLIP favors the new prompt over the old one. Please also show a measured trace of the gate weight around a prompt switch to back the recovery claim in App. B, and clarify whether the descriptor caches are cleared at sequence boundaries, as this is not explicitly specified in Alg. 1.
2 comments
AuthorsOpenReview ↗

Thank you for your time and feedback - happy to discuss further

Dear Reviewer h1iz,

Thank you very much for your time and thoughtful feedback on our work. As the discussion period is nearing its end, we would be grateful if you could take a look at our rebuttal when you have a chance. Please let us know if there is anything we can clarify or elaborate on; we would be happy to discuss further.

Thank you again for your time and consideration.

Best regards, Authors of Submission #3673

Thank you for the detailed rebuttal and the revised manuscript.

My first remaining concern on W1 is the evidence that the drift originates in the teacher. As I understand it, the claim of the paper is that the teacher output is prompt-consistent yet trajectory-inconsistent. The counterfactual experiment shows that a shift arises on the teacher side even when the student is unchanged. However, a change in the teacher output under a different prompt is expected by construction, so it remains unclear to me whether this shift is trajectory-inconsistent. It would be helpful if you could show how much of the change in the output is requested by the prompt and how much is not. For example, one could measure how much the subject and background that are described in both prompts differ between the two teacher outputs.

My second remaining concern on W1 is that the student displacement is contained in the gate signal. The Fast 25% analysis confirms that the gate decreases less than at a prompt switch. Looking at the numbers, however, it seems difficult to say that the gate is unchanged. A statistical test on the difference between Steady and Fast 25% would be helpful. Also, as I understand it, the Switch Gate is measured at chunks where both the prompt and the student chunk change. It would be helpful to see, as a control, the gate weight under the condition where the student is fixed and only the prompt is changed (Same (x, ε)).

Regarding Q1 and Q2, the ablation experiments have been extended well. I would like to ask for one confirmation. Table 7 is reported at 1,000 steps, whereas the experimental setting in the main text describes 3,000 steps. Was the setting changed for the ablation table? The numbers for Ours appear to be identical to those in Table 3. My other concerns have been well addressed.

As my concerns on W1 remain, I will keep my score for now. Thank you again for your thoughtful responses.

Reviewer_RU91OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Summary

This paper addresses trajectory drift in interactive autoregressive video generation. It attributes part of this problem to conditional bias in the teacher’s guidance and proposes using temporal changes in the teacher’s predictions to adaptively weight distribution-matching supervision.

Critical strengths
  1. Important problem. Maintaining temporal consistency is a significant challenge in streaming video generation.

  2. Valuable perspective. Investigating the teacher as a potential source of drift offers a useful perspective beyond explanations centered on error accumulation in the student.

Critical weaknesses
  1. The reliability criterion may not sufficiently distinguish drift from legitimate changes. The diagnosis of conditional bias is plausible: the teacher’s short window guidance may be incompatible with the student’s long history. However, the proposed method relies on unusually large changes in DINO features. A legitimate abrupt event, such as a door suddenly opening, could trigger the detector, whereas harmful gradual drift could go undetected.

  2. The method introduces several hyperparameters and additional training complexity. These include the gate threshold z0z_0, gate sharpness γz\gamma_z, continuity weight λcont\lambda_{\mathrm{cont}} and probe noise level. Although the paper provides sensitivity analyses for some parameters, the robustness of the overall configuration remains unclear. I wonder if a simpler alternative, such as a standard streaming autoregressive framework like Astrolabe augmented with a VLM-based consistency judge could work.

Questions for authors
  1. Could the consistency gains partly result from suppressing motion? By downweighting teacher predictions associated with unusually large feature changes, the method may also discourage legitimate rapid motion. A nearly static video can exhibit little drift while offering limited value for interactive generation. Meanwhile, Table 2 also reported the degenerated dynamics, beaten by the Reward-Forcing, which strengthen my concern.

  2. How does the Causal Forcing initialization perform? The method uses Causal Forcing for initialization, but it is not reported as a separate baseline in the main tables. Could the authors report its performance?

2 comments
AuthorsOpenReview ↗

Thank you for your time and feedback - happy to discuss further

Dear Reviewer RU91,

Thank you very much for your time and thoughtful feedback on our work. As the discussion period is nearing its end, we would be grateful if you could take a look at our rebuttal when you have a chance. Please let us know if there is anything we can clarify or elaborate on; we would be happy to discuss further.

Thank you again for your time and consideration.

Best regards, Authors of Submission #3673

Thanks for you rebuttal. My concern regarding the Causal-forcing initialization and motion compression has been addressed. However, I remain concerned about the claim that “sudden events are not frozen.” You explain that Eq. 5 detects standardized teacher self-inconsistency outliers rather than motion speed. Yet a legitimate abrupt event can also produce such an outlier. For example, a closed door suddenly opening reveals a previously occluded scene, potentially causing a substantial feature change relative to the preceding history. You mentioned "the loss penalizes 2nd-order acceleration". However, a door beginning to open must transition from rest to motion, which is exactly acceleration. Similar examples include people first stand and still then turn around. you would also punish this And your table omits Long-CLIP and VideoAlign results for the hyperparameter sweeps. Can you add them? I would keep my score for now.

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

This paper investigates long-term temporal consistency in interactive autoregressive video generation, focusing on the challenge of responding to evolving event conditions while preserving previously established visual trajectories.

The authors identify conditional bias as a potential source of cross-event trajectory drift in existing distribution-matching distillation (DMD) pipelines. Specifically, a frozen bidirectional diffusion teacher may generate semantically plausible guidance under updated prompts that is nevertheless inconsistent with the student's previously generated history.

To address this issue, the paper proposes Delta Forcing, which consists of two main components:

Student Momentum Preservation, which regularizes changes in consecutive feature-space displacements to encourage continuous trajectory evolution.

History-Calibrated Teacher Reliability, which estimates teacher reliability using historical statistics of feature displacements and adaptively balances DMD supervision with student trajectory continuity regularization.

Experiments on the MemFlow benchmark show improvements in subject and background consistency, along with competitive prompt alignment and human preference results.

Critical strengths

Relevant and well-motivated problem. The tension between responsiveness to evolving prompts and long-term temporal consistency is an important challenge in interactive autoregressive video generation. The paper focuses on a practical failure mode that is not fully addressed by existing streaming generation methods.

Intuitive and practical method. The proposed combination of adaptive teacher supervision and student trajectory regularization is conceptually simple and can be incorporated into an existing distribution-matching training pipeline.

Encouraging empirical results. Delta Forcing achieves improvements in subject and background consistency over several competitive baselines. The qualitative examples further illustrate reduced appearance and layout drift across event transitions.

Human evaluation and useful sensitivity analyses. The user study provides evidence beyond automated metrics, while the hyperparameter and structural ablations help characterize the trade-off between consistency and dynamic behavior.

Critical weaknesses
  1. Insufficient evidence supporting conditional bias as the primary cause of trajectory drift.

A central contribution of the paper is the identification of teacher-induced conditional bias as a fundamental source of cross-event trajectory drift. However, the evidence supporting this diagnosis is primarily qualitative, relying on selected teacher-student visualizations and latent trajectory projections.

These observations suggest a correlation between teacher transitions and student drift, but do not conclusively establish that the teacher's conditional bias is the primary cause, rather than other factors such as limited context, student rollout errors, or event-switching mechanisms.

Moreover, the reliability gate in Eq. (5) is determined solely by the standardized displacement of consecutive teacher features, rather than by directly measuring the disagreement between teacher-induced transitions and the student's history-grounded trajectory. This raises the question of whether the proposed mechanism actually detects unreliable teacher guidance or simply suppresses unusually large temporal changes.

A quantitative analysis of the relationship between gate activations, teacher-student inconsistencies, and observed drift would provide stronger evidence for the central hypothesis.

  1. Potential suppression of legitimate event transitions.

The proposed reliability mechanism assumes that unusually large teacher feature displacements indicate unreliable guidance. However, in interactive video generation, some legitimate prompt changes may require substantial and immediate visual transitions, such as the sudden appearance of new objects, abrupt action changes, or rapid camera movements.

Such transitions may also produce large feature displacements and trigger the gate, potentially suppressing valid supervision and delaying the model's response to new instructions.

Although Appendix B acknowledges this limitation, the paper lacks controlled experiments distinguishing undesirable trajectory drift from legitimate abrupt changes.

The reported Long-CLIP scores provide evidence of general prompt alignment, but may not adequately capture response latency or the correctness of event execution immediately after a prompt switch.

  1. Insufficient quantitative ablations and comparisons with simpler alternatives.

The paper provides useful hyperparameter sensitivity analyses, but the individual contributions of the two main components are primarily demonstrated through qualitative examples in Fig. 4. Quantitative results for removing the continuity loss and adaptive trust-region mechanism would make their respective contributions clearer.

More importantly, the proposed gate should be compared with simpler strategies, such as uniformly reducing the DMD weight around prompt-switching boundaries or applying a fixed temporal continuity regularizer.

Since the identified failure mode is strongly associated with event transitions, a simple event-boundary-aware baseline might already mitigate part of the observed drift. Without such comparisons, it is difficult to determine whether the gains specifically arise from the proposed reliability estimation mechanism.

  1. Limited evaluation of generalization across interactive scenarios.

The experiments are primarily conducted using a single generator configuration and the MemFlow benchmark, consisting of 60-second videos with six consecutive 10-second events.

As described in Appendix E, the benchmark emphasizes continuous scenes in which persistent entities are repeatedly specified and events evolve incrementally. While this is a relevant setting, it may not fully represent the diversity of real-world interactive generation.

It remains unclear how the method performs under more abrupt or semantically diverse changes, different prompt-switching frequencies, longer generation horizons, or alternative backbone architectures.

A broader evaluation would help substantiate the generality of the proposed method.

  1. Limited statistical analysis of the human evaluation.

The inclusion of a human study is valuable. However, the evaluation involves only 20 participants, each completing four ranking trials, and the paper reports average ranks without uncertainty estimates or statistical significance tests.

Additionally, forced ranking does not allow participants to indicate that two methods are perceptually indistinguishable. Consequently, the reported ranking improvements do not directly establish the frequency or magnitude of perceptually meaningful differences.

Reporting confidence intervals, statistical significance, inter-rater agreement, and the number of unique evaluated sequences would strengthen the human evaluation.

Questions for authors

Q1. Validation of Conditional Bias. Can the authors provide quantitative evidence that the proposed gate identifies teacher-induced trajectory inconsistencies rather than general temporal changes? In particular, how do the gate values correlate with actual teacher-student disagreement and observed appearance or layout drift around event boundaries?

Q2. Legitimate Abrupt Transitions. How does Delta Forcing handle prompt changes that require immediate and substantial visual changes? Could the authors evaluate abrupt object appearances, sudden action changes, and camera movements, and report both transition consistency and response latency?

Q3. Simpler Baselines and Component Ablations. Could the authors provide quantitative results for removing Design 1 and Design 2 separately? Additionally, how does Delta Forcing compare with a simpler event-boundary-aware strategy that reduces DMD supervision for a fixed number of chunks after each prompt switch?

Q4. Generalization. Does the proposed approach maintain its advantages under different event-switching frequencies, longer generation horizons, and alternative autoregressive video generation architectures?

Q5. Human Evaluation. Could the authors report confidence intervals or statistical significance tests for the user-study results, along with the number of unique sequences evaluated? A pairwise human evaluation allowing ties would also help establish whether the observed preference improvements are perceptually meaningful.

1 comment
AuthorsOpenReview ↗

Thank you for your time and feedback - happy to discuss further

Dear Reviewer SzKE,

Thank you very much for your time and thoughtful feedback on our work. As the discussion period is nearing its end, we would be grateful if you could take a look at our rebuttal when you have a chance. Please let us know if there is anything we can clarify or elaborate on; we would be happy to discuss further.

Thank you again for your time and consideration.

Best regards, Authors of Submission #3673

Comments on the paper

9 comments

Q2: "Legitimate Abrupt Transitions. How does Delta Forcing handle prompt changes that require immediate and substantial visual changes? Could the authors evaluate abrupt object appearances, sudden action changes, and camera movements, and report both transition consistency and response latency?" (Also addressing W2 on potential suppression or delay of valid transitions.)

Delta Forcing executes sudden legitimate transitions without delay for two reasons: (1) outlier clipping (μk+4σk\mu _ k + 4\sigma _ k) prevents switch spikes from inflating running statistics (μk,σk)(\mu _ k, \sigma _ k), restricting gate dips to 1–2 chunks (0.75–1.5s0.75\text{–}1.5\text{s}) before wkw _ k immediately rebounds (0.85–0.93) to full DMD supervision under PnewP _ {\text{new}}; and (2) Design 1 (Lcont=∥δkF−sg(δk−1F)∥22\mathcal{L} _ {\text{cont}} = \Vert\boldsymbol{\delta}^F _ k - \text{sg}(\boldsymbol{\delta}^F _ {k-1})\Vert _ 2^2, Eq. 3) regularizes second-order acceleration rather than first-order frame similarity, allowing immediate velocity changes for new actions or camera moves while preventing zero-chunk scene resets.

Notation: Beyond Subj. / Bg. (defined in Q1), Table R3 reports:

  • Smooth. / Dyn. and Aesth. / Img.: VBench Motion Smoothness / Dynamic Degree (%) and Aesthetic / Imaging Quality; LC Avg.: 10s segment-level Long-CLIP alignment averaged over S1–S6S _ 1\text{–}S _ 6 (as in Table 1).
  • Onset Latency (Obj / Act / Cam): chunk-level (0.75s/chunk) post-switch time until chunk-wise Long-CLIP margin ΔS(k)=S(ck,Pnew)−S(ck,Pold)\Delta S(k) = S(c _ k, P _ {\text{new}}) - S(c _ k, P _ {\text{old}}) rises above the 4-chunk pre-switch baseline (ΔS(k)>max⁡(0,ΔSpre+0.005)\Delta S(k) > \max(0, \Delta S _ {\text{pre}} + 0.005)), reported overall and for sudden object appearances (Obj), action changes (Act), and camera moves (Cam).
  • Early 2s CLIP: Long-CLIP alignment with PnewP _ {\text{new}} averaged strictly over the immediate post-switch window [0, 2.25s] (first 3 chunks).

Experiment — Stress-testing abrupt transitions and immediate post-switch latency (Table R3): To prove that Delta Forcing preserves fast prompt execution alongside transition consistency on sudden visual changes, we constructed a 60-case Abrupt Benchmark (Abrupt 60: 60 60s 6-event clips =300= 300 switches, split evenly into 20 object appearances, 20 sudden action changes, and 20 rapid camera movements) and evaluated it alongside all 500 switches of Main 100 (Tables 1–2). Why this setting fits our scenario: Whereas 10s segment-averaged Long-CLIP could mask a 1–2s early lag (as W2 noted), Abrupt 60 directly stress-tests the exact three abrupt transition types in Q2, while 0.75s-resolution Onset Latency and Early 2s CLIP ([0, 2.25s]) pinpoint immediate post-switch responsiveness right where the gate activates.

Table R3: Prompt Responsiveness & VBench on 60-Case Abrupt and 100-Case Main Benchmarks

Benchmark & MethodSubj. / Bg. ↑Smooth. / Dyn. ↑Aesth. / Img. ↑LC Avg. ↑Onset Latency ↓ (Obj / Act / Cam)Early 2s CLIP ↑
Abrupt 60: LongLive93.70 / 93.0097.49 / 94.1856.10 / 72.2125.952.74s (2.85 / 2.66 / 2.70s)25.58
Abrupt 60: Delta Forcing95.08 / 94.1098.16 / 95.4256.52 / 72.8926.282.80s (2.96 / 2.59 / 2.85s)26.12 (+0.54)
Main 100: LongLive94.97 / 93.3798.37 / 90.9358.95 / 72.5926.055.32s25.41
Main 100: Delta Forcing96.63 / 94.8298.71 / 92.3258.95 / 73.1126.245.47s (+0.15s)25.98 (+0.57)
  • Virtually zero response delay across all abrupt categories (addressing Q2 latency): Delta Forcing matches un-gated LongLive in onset latency within 0.06s on Abrupt 60 (2.80s vs. 2.74s, <0.1<0.1 chunk—and is 0.07s faster on sudden actions, 2.59s vs. 2.66s) and within 0.15s on Main 100 (5.47s vs. 5.32s, 0.2 chunk), confirming the brief 1–2 chunk gate dip does not delay new instructions.
  • Simultaneously higher early alignment, motion, and transition consistency (addressing Q2 consistency & W2): Instead of suppressing valid transitions, Delta Forcing improves immediate post-switch Early 2s CLIP by +0.54 / +0.57 (26.12 vs. 25.58; 25.98 vs. 25.41), Dynamic Degree by +1.24 / +1.39 (95.42% / 92.32%), and Subject / Background Consistency by +1.38 / +1.10 on Abrupt 60 (+1.66 / +1.45 on Main 100).
  • Qualitative & demo video note: We also provide full-length demo videos illustrating such abrupt transitions—for instance, in Fig. 3 Case 1 (prompts in Appendix E, Example@1), the newly prompted green dragon promptly flies in and lands on the wizard's shoulder in the 10–20s segment while preserving the wizard's identity and chamber layout. According to the policy, we attach an anonymous project link for demo: https://anonymous-rebuttal-delta-forcing.vercel.app/#

Q5: "Human Evaluation. Could the authors report confidence intervals or statistical significance tests for the user-study results, along with the number of unique sequences evaluated? A pairwise human evaluation allowing ties would also help establish whether the observed preference improvements are perceptually meaningful." (Also addressing W5 and Ethics.)

We candidly note that we did not run a tie-allowing Good/Same/Bad (GSB) study in the submission; we chose simultaneous 4-way ranking in Table 3 to avoid central-tendency ("Same") fatigue over 60s videos, and we will do our best to add a tie-allowing pairwise GSB evaluation in the revision.

Our Table 3 preference gains are highly significant with substantial inter-rater agreement across 50 unique 60s 6-event sequences (200 generated 60s videos across 4 models). Specifically, 20 domain-informed raters each completed 4 blind, randomized 4-way ranking trials (204=80 trials: 30 unique sequences independently double-rated by two raters to measure inter-rater concordance, plus 20 single-rated sequences to cover half of the 100-case benchmark).

Notation: In Table R6:

  • Mean Rank [95% CI] (, 1 = Best, 4 = Worst): mean rank across the 80 blind 4-way trials with 10,000-resample non-parametric bootstrap 95% CIs.
  • (p) in each baseline column: Holm-Bonferroni corrected paired Wilcoxon signed-rank p-value testing Delta Forcing vs. that baseline.
  • Friedman 2 (p): omnibus non-parametric Friedman statistic and p-value across all 4 simultaneously ranked methods.
  • Kendall's W[0,1]: inter-rater coefficient of concordance on the 30 independently double-rated 60s sequences (W>0.6 denotes substantial agreement).

Statistical Analysis — Non-parametric uncertainty, significance & concordance testing (Table R6): We conducted full non-parametric testing over all 80 blind 4-way trials (50 unique 60s sequences) from Table 3 (Sec. 5.2 & App. F). Because comparative ranks (1–4) are bounded ordinal variables rather than Gaussian scores, bootstrap 95% CIs, omnibus Friedman 2, and Holm-corrected paired Wilcoxon signed-rank tests provide exact distribution-free inference while controlling family-wise Type-I error; meanwhile, our hybrid 30 × 2 double-rated + 20 × 1 single-rated design jointly quantifies inter-rater reliability and sequence diversity.

Table R6: User Study Ranks (1=Best, 4=Worst) [95% CI], Holm-Wilcoxon p, Friedman 2 & Kendall's W (50 Unique Sequences)

Criterion (Table 3)Delta Forcing ↓LongLive ↓ (p)MemFlow ↓ (p)Reward Forcing ↓ (p)Friedman 2 (p)Kendall's W
Aesthetic Quality2.13 [1.90, 2.37]2.78 [2.51, 3.04] (3.4e-3)2.68 [2.42, 2.94] (1.0e-2)2.41 [2.14, 2.68] (1.1e-1)12.43 (6.0e-3)0.600
Dynamic Quality1.93 [1.69, 2.19]2.44 [2.18, 2.70] (1.0e-2)3.08 [2.86, 3.29] (2.6e-7)2.55 [2.28, 2.82] (4.0e-3)32.74 (3.7e-7)0.653
Multi-Event Naturalness1.74 [1.55, 1.95]2.63 [2.38, 2.89] (1.6e-6)2.89 [2.62, 3.16] (2.9e-7)2.74 [2.51, 2.96] (1.8e-6)38.73 (2.0e-8)0.610
Average Rank1.93 [1.78, 2.09]2.62 [2.44, 2.79] (1.9e-6)2.88 [2.70, 3.06] (4.1e-9)2.57 [2.39, 2.76] (1.6e-5)42.65 (2.9e-9)0.636
  • Significant preference & strong rater agreement (Q5 part 1, W5): Delta Forcing ranks first on every criterion, achieving non-overlapping 95% CIs against all baselines on Multi-Event Naturalness (1.74 [1.55, 1.95] vs. 2.63–2.89, all pHolm1.810-6, Friedman p=2.010-8) and Average Rank (1.93 [1.78, 2.09] vs. 2.57–2.88, all pHolm1.610-5, Friedman p=2.910-9), plus significant Dynamic Quality gains (pHolm1.010-2). Crucially, Kendall's W=0.600–0.653 (overall 0.636) confirms consistent rater agreement on cross-event continuity, ruling out random tie-breaking on indistinguishable clips.
  • Ethics Review note: As detailed in our Ethics Statement (p. 10), our study involved only anonymous adult volunteers rating non-sensitive synthetic clips (landscapes, animals, everyday scenes) with informed consent, zero PII, and minimal risk under institutional IRB-exempt guidelines.

Experiment 2 — Optical-flow decomposition & comparison with Reward Forcing (Table R7): Sequence-wide VBench Dynamic Degree conflates genuine motion with two artifacts: switch scene-reset jumps (inflating 0–1.5s flow) and RL mode-seeking camera panning. We decompose Farneback optical flow across all 500 switches of Main 100 and 300 switches of Abrupt 60 into Boundary Flow (0–1.5s) and Steady Flow (1.5–10s) to cleanly separate genuine motion from boundary resets and drift.

Table R7: Switch-Boundary vs. Steady-State Optical Flow & Motion Quality (Main 100 & Abrupt 60)

Benchmark & MethodSubj. / Bg. ↑Smooth. / Dyn. ↑VideoAlign MQ ↑Boundary (0–1.5s)Steady (1.5–10s)Bnd / Steady
Main (100, 500 sw): Stage-1 Init96.19 / 94.7598.45 / 76.40—1.171.240.95×
Main (100, 500 sw): LongLive94.97 / 93.3798.37 / 90.933.961.681.211.39×
Main (100, 500 sw): Delta Forcing96.63 / 94.8298.71 / 92.324.191.181.240.96×
Abrupt (60, 300 sw): LongLive93.70 / 93.0097.49 / 94.18—2.862.281.25×
Abrupt (60, 300 sw): Delta Forcing95.08 / 94.1098.16 / 95.42—2.292.350.97×
  • Higher genuine motion without boundary reset spikes: LongLive's un-gated switch bias causes abrupt scene resets, creating artificial 1.39× (Main) and 1.25× (Abrupt) boundary flow spikes (1.68 vs. 1.21; 2.86 vs. 2.28). Delta Forcing eliminates these spikes (0.96× / 0.97×) while achieving higher genuine Steady Flow (1.24 vs. 1.21; 2.35 vs. 2.28) and higher VBench Dynamic Degree (92.32% vs. 90.93% on Main; 95.42% vs. 94.18% on Abrupt 60 with 0.06s onset parity, SzKE Q2 Table R3). Static collapse occurs only under a non-adaptive Constant Gate (wk0.5: 41.32% Dyn., SzKE Q3 Table R4).
  • Why Reward Forcing's VBench Dyn. (93.16) is drift-inflated—and why Delta Forcing wins on human Dynamic Quality: Raw VBench Dynamic Degree rewards any pixel displacement, including unintended DMD camera drift (Fig. 4 row 4) and Reward Forcing's RL mode-seeking camera panning (Sec. 6, Fig. 5 Right)—which inflates its raw flow to 93.16 while collapsing prompt alignment (Long-CLIP 24.75 vs. 26.24, VideoAlign Text Alignment 0.04 vs. 0.57, Table 1). When evaluated on meaningful, prompt-faithful motion, Delta Forcing ranks #1 in VideoAlign MQ (4.19, Table 1) and #1 in our blind User Study on Dynamic Quality (1.93, Table 3 & App. F), significantly beating Reward Forcing (2.55, pHolm=4.010-3), LongLive (2.44, pHolm=1.010-2), and MemFlow (3.08, pHolm=2.610-7, Kendall's W=0.653).
  • Qualitative verification on sudden events: In Fig. 3 Case 1 (App. E, Example@1), when the 10–20s prompt suddenly introduces a green dragon flying in through a window onto the wizard's shoulder, Delta Forcing executes the rapid entry promptly and fluidly while keeping the wizard and chamber intact. Accoding to the policy, I attach an anonymous project link for demo: https://anonymous-rebuttal-delta-forcing.vercel.app/#

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.