PaperlayerView exact revision on arXiv

Source-linked AI summary

SNF-Bench: Separating Static Drift from Natural Flow in Long-Horizon Fixed-Camera Video Generation

Matiur Rahman Minar, Seunghun Oh, Ganghyeon Jeong, Unsang Park

arXiv:2608.28694v1cs.CV

TL;DR

Long-horizon fixed-camera evaluation can conflate desirable natural flow with background drift. SNF-Bench separates static support from dynamic flow, and its audit finds that whole-frame motion and static-region drift can yield near-opposite interpretations of the same outputs.

  • Problem

    Long-horizon fixed-camera nature-video evaluation lacks measures separating desired dynamic flow from background drift, while whole-frame metrics conflate distinct errors.

    Source: §1 · p. 1

  • Method

    SNF-Bench partitions scenes into static support and dynamic flow, reporting static fidelity and flow persistence with absolute magnitude separately, with drift leakage as context.

    Source: §1 · p. 2, §4 · p. 4

  • Results

    Whole-frame motion and static-region drift yield near-opposite interpretations of the same outputs; at maximum translation, fBD and NBF reach 1.86× and 1.32× baseline, while VBench Dynamic Degree reaches only 1.07×.

    Source: §1 · p. 2, §5.2 · p. 6

  • Takeaways & Limitations

    Whole-frame motion and region-resolved static fidelity can materially change interpretation of the same outputs, complementing broad benchmarks without a composite score.

    Source: §8 · p. 8

    SNF-Bench is limited to long-horizon fixed-camera generation where static support and dynamic flow separate, and does not measure physical or semantic realism.

    Source: §8 · p. 8

Abstract

from arXiv · show

Long-horizon video generation is evaluated with whole-frame metrics that reward motion and temporal consistency. For fixed-camera nature scenes this creates an ambiguity: motion of water, fire, smoke, or rain is desirable, whereas motion of the background is an error. A system can therefore score well on motion while its scene drifts, or on consistency while its flow stagnates. We introduce SNF-Bench, an evaluation framework for long-horizon fixed-camera generation that partitions each scene into static support and dynamic flow and reports static fidelity, flow persistence with absolute magnitude, and drift leakage separately, never as one score. Drift leakage is interpretive context rather than a headline measurement. Each factor is validated mechanistically rather than by correlation with preference: we inject global translation, rotation, and scale drift and progressive late freezing at known severity into real generations, and require each factor to respond in its stated direction and to remain selective against corruptions it does not target. Auditing publicly released long-horizon text-conditioned checkpoints under one recorded common inference configuration, plus an image-conditioned track with released-pipeline references and a deployment-sensitivity panel, we find that whole-frame motion and static-region drift induce near-opposite orderings of the same outputs. At maximum controlled translation, fBD and NBF rise to $1.86\times$ and $1.32\times$ baseline, but whole-frame Dynamic Degree reaches only $1.07\times$---rewarding the corruption. SNF-Bench measures where motion occurs and whether it persists; it does not measure physical realism. Project page: https://minar09.github.io/snfbench/.

1. Introduction · 2. Related Work

SNF-Bench separates static-scene fidelity from persistent natural flow because whole-frame metrics can conflate background drift with desired motion or reward stagnant flow. Related work motivates a complementary, mechanistically validated evaluation focused on where motion occurs and whether it persists.

1. Introduction

  • Whole-frame metrics can conflate accumulated static-background drift with desired motion, while consistency measures can reward suppressing flow in fixed-camera nature scenes.The distinction applies to scenes such as rivers, waterfalls, waves, rain, fire, smoke, and wind-driven vegetation.Source: §1 · p. 1
  • Known-severity perturbations test whether each factor responds in its defined direction and remains selective, with geometric severity expressed as induced mean displacement.The protocol covers translation, rotation, scale, motion attenuation, late freezing, photometric drift, temporal repetition, and region-partition erosion and dilation.Source: §1 · p. 2
  • SNF-Bench’s audit finds that whole-frame metrics and spatially decomposed measures can interpret the same long-horizon outputs differently, without claiming existing metrics are incorrect or measuring physical realism.The framework makes the spatial origin and temporal persistence of motion explicit; drift leakage is interpretive context, not a headline measurement.Source: §1 · p. 2
  • SNF-Bench partitions static support from dynamic flow and reports static fidelity, flow persistence with absolute compensated magnitude, and drift leakage separately rather than collapsing them into one score.The benchmark evaluates primarily at 60 and 120 s, retaining 5 s and 240 s tiers as diagnostics.Source: §1 · pp. 1–2

2. Related Work

  • General-purpose suites such as VBench [12], EvalCrafter [21], and Video-Bench [7] evaluate broad video qualities, but aggregate measures do not identify whether motion comes from intended flow or displaced static structure.VMBench [19] is the closest comparator, assessing perceptual plausibility, whereas SNF-Bench asks where motion occurs and whether it survives compensation for global displacement; its mechanistic validation complements human-alignment evidence.Source: §2 · pp. 2–3
  • Long-horizon autoregressive and streaming generators, including CausVid [36], LongLive [33], and Causal-Forcing [42], differ in temporal block size, training objective, memory/cache policy, and horizon-extension mechanism, yet evaluations typically use generic video metrics or method-specific qualitative analysis.This diversity motivates a recorded common-configuration audit, while the evaluation gap motivates asking whether static support remains fixed while dynamic flow persists.Source: §2 · p. 3
  • Image-animation research separates static structure from natural motion, including controllable fluid animation [24] and Generative Image Dynamics [16]; SNF-Bench makes that distinction an evaluation principle for long-horizon generators.Source: §2 · p. 3

3. SNF-Bench Design · 4. Drift–Motion Evaluation Protocol

SNF-Bench evaluates fixed-camera long-horizon videos by separating stationary support from intended natural flow, measuring where motion occurs and whether it persists without reducing the factors to one score. Its protocol pairs static-region fidelity with early and late dynamic-flow magnitudes, while treating drift diagnostics as context rather than causal or semantic judgments.

3. SNF-Bench Design

  • SNF-Bench partitions fixed-camera scenes into stationary support and persistent natural-flow regions, evaluating T2V and I2V under shared conditioning and decomposition.The partition defines evaluation roles, not semantic classes, and does not require flow to follow a physical model.Source: §3 · p. 3

3.1. Scene Categories and Horizons; 3.3. Benchmark Manifest

  • The benchmark fixes natural-process categories before evaluation, reports prompt means at 60 s and 120 s, and records unsupported target horizons rather than silently omitting them.Channel water makes up 61% of the principal prompt set, so aggregates are not category macro-averages; a versioned manifest freezes items and prompts before the sweep.Source: §3.1 · p. 3, §3.3 · p. 4

3.2. Region Partition

  • An automatic early-window flow partition assigns dynamic pixels by Otsu thresholding and excludes eroded static boundaries and a 4% frame border from measurement.At 60 s, the partition uses the first 7.1 s to limit circularity from late drift, but early drift can still affect it; boundary erosion and dilation sensitivity are tested, and precipitation static-fidelity values include unmodeled transit motion.Source: §3.2 · pp. 3–4

4. Drift–Motion Evaluation Protocol; 4.4. Joint Interpretation

  • The protocol reports static fidelity jointly with relative flow persistence and absolute dynamic-motion magnitude; drift leakage contextualizes these axes, and no single score is sufficient.Low drift can reflect freezing, while high foreground motion can accompany whole-scene drift; videos are subsampled without interpolation and globally fixed estimators and settings are applied across systems.Source: §4 · p. 4, §4.4 · p. 5

4.1. Static Fidelity

  • fBD measures late-window displacement of matched static-region features, while NBF measures static-mask flow magnitude normalized by frame width and elapsed time.NBF is reported in 10^-3 frame widths per second, making it comparable across native frame rates and sampling policies; it is a magnitude, not a ratio to whole-frame flow.Source: §4.1 · p. 4

4.2. Flow Persistence

  • Global similarity compensation removes estimated static-scene translation, rotation, and scale before MCFF is measured, and MCFF is reported in both early and late windows.Flow Persistence is late-to-early MCFF, clipped to prevent instability when early motion is small; it is never reported without MCFF magnitudes, and values above one are not automatically better.Source: §4.2 · p. 4

4.3. Drift Leakage and Attenuation

  • DLR compares late static-region flow with raw dynamic-region flow, while DAR measures relative attenuation after compensation; neither is a causal decomposition of motion.DLR’s rotational response is inverted, so it is read alongside fBD and DAR; SNF-Bench also does not establish that the depicted phenomenon matches the prompt and should be paired with a semantic score.Source: §4.3 · p. 5

5. Controlled Validation

Controlled corruptions test whether SNF-Bench factors respond in the expected direction and selectively to their target failure modes. Results show that translation and freezing are detected, while rotation and repeated content expose important limits.

5. Controlled Validation

  • Validation uses known corruption types and severities, with perturbation strengths, masks, windows, and hyperparameters fixed before final results are inspected and independently of the Sec. 6 audit.Admission requires target-family Spearman correlation ρ ≥0.7 and off-target response below 25%; rotation is reported but not gated.Source: §5 · p. 5, Table 1 · p. 6

5.1. Perturbations and Criteria

  • Corruptions isolate geometric drift, motion attenuation, late freezing, photometric drift, temporal repetition, and partition changes, with geometric severity normalized as mean pixel displacement.This normalization makes severity comparable across families whose native parameters differ substantially; factors must respond monotonically in the expected direction, remain selective to unrelated corruptions, and survive partition perturbation.Source: §5.1 · pp. 5–6, Table 1 · p. 6

5.2. Validation Results

  • At maximum translation, fBD and NBF reach 1.86× and 1.32× baseline, while VBench Dynamic Degree [12] reaches only 1.07×; late freezing drives MCFF-L and FP down to 0.02 and 0.07 of unperturbed values.Freezing leaves static factors comparatively unchanged, as required by the decomposition.Source: §5.2 · p. 6
  • Rotation weakens both static-fidelity factors and inverts DLR; looping sustains persistence and magnitude, while precipitation exclusion preserves NBF ordering and the area–fBD association remains an open validity concern.Under rotation, fBD and NBF respond +0.70 and +0.60, versus +1.00 for both under translation; DLR responds −0.80. SNF-Bench does not distinguish genuine progression from repetition.Source: §5.2 · p. 6, Table 2 · p. 6

6. Public-Model Audit · 7. Results and Interpretation Changes

The public-model audit evaluates checkpoints under a recorded common configuration and reports static fidelity and flow behavior separately. Its results show that whole-frame motion can tell a different story from static-region flow and late-motion persistence.

6. Public-Model Audit

  • The audit covers seven publicly released generators under one recorded T2V configuration, not each checkpoint’s published inference procedure.The shared setup uses four denoising steps, six frames per block, seed 0, and 832 × 480 output at 16 fps; distilled checkpoints are not tested at their own step budgets.Source: §6 · p. 6
  • The audit fixes conditioning and seed within each item, averages per-item scores, and estimates uncertainty with percentile bootstrap intervals over items.Paired-item design supports cross-system comparisons; the reported focus is interpretation change, not a claim that existing whole-frame metrics are wrong.Source: §6 · pp. 6–7

7. Results and Interpretation Changes

  • T2V comparisons refer to checkpoints under the recorded common configuration, while released-pipeline and wrapper I2V results are identified separately and never ranked across settings.The I2V panels show deployment sensitivity without attributing its cause.Source: §7 · p. 7, §7.3 · p. 8

7.1. Recorded Common-Configuration T2V Audit

  • Sustained flow requires both non-trivial late motion and high retention; MCFF-E alongside MCFF-L exposes uniformly negligible motion.Table 3 reports MCFF-E→L with FP, whose per-video clipped-ratio mean need not equal the ratio of the aggregate MCFF means.Source: §7.1 · p. 7, Table 3 · p. 7
  • Under the recorded common configuration, Infinite-Forcing has the lowest NBF but the least surviving late motion, while Causal-Forcing [42] occupies the opposite, higher-background-flow corner.Paired per-prompt intervals for Infinite-Forcing’s NBF versus all six peers exclude zero.Source: Table 3 · p. 7, §7.1 · p. 7

7.2. Whole-Frame Motion and Static-Region Flow Yield Different Interpretations

  • Across seven checkpoint means, Dynamic Degree tracks static-region flow at descriptive ρ = +0.93; Causal-Forcing ranks 1/7 on Dynamic Degree but 7/7 on NBF, with Infinite-Forcing reversed.The per-item analysis is reported in the supplementary material.Source: §7.2 · p. 7
  • Causal-Forcing’s DLR of 0.94 indicates static-region motion energy nearly equal to dynamic-region energy, while its DAR of 0.40 indicates attenuation from fitted similarity compensation.The residual may reflect non-rigid instability, intended motion, partition contamination, or flow error; DAR is not a causal decomposition.Source: §7.2 · pp. 7–8

7.3. Image-Conditioned (I2V) Track: Released Pipelines and Deployment Sensitivity

  • At 60 and 120 seconds, the I2V track reports two released-pipeline rows and five dagger-marked wrapper rows per horizon, keeping the settings separate.The 60-second figure uses a shared source image, so static support is given rather than imagined differently.Source: Table 4 · p. 8, §7.3 · p. 8, Figure 4 · p. 8

8. Limitations and Conclusion

SNF-Bench separates static fidelity and flow persistence to clarify interpretations that whole-frame motion can obscure, without combining them into a composite score. Its scope and automatic partition impose limits, while the released specification and common-configuration audit support reproducible evaluation.

Limitations

  • SNF-Bench targets long-horizon fixed-camera scenes where static support and dynamic flow separate, but does not measure physical or semantic realism.It cannot distinguish progression from repetition and relies on an automatic partition bounded by an erosion–dilation study.Source: §8 · p. 8

Conclusion

  • Separating static fidelity from flow persistence can change how the same outputs are interpreted, complementing broad benchmarks without a composite score.Drift quantities are reported as context rather than as headline measurements.Source: §8 · p. 8
  • The released specification and common-configuration checkpoint audit provide a reproducible evaluation basis, while independent partitions and broader non-water coverage remain next steps.Source: §8 · p. 8

Supplementary Material · S1. Formal Factor Definitions

The supplement collects supporting procedures, audits, coverage, perceptual checks, and qualitative material for both tracks. S1 defines the notation and measurement conventions for background drift and flow, motion compensation, foreground-flow persistence, and drift leakage and attenuation.

Supplementary Material

  • The supplement reports both tracks at equal depth and collects region-partition sensitivity, complete audit tables, measurement coverage, perceptual-check protocols, and qualitative material.It also states that benchmark code and data are released.Source: Supplementary Material · p. 11

S1. Formal Factor Definitions

  • S1 defines the sequence, region masks, optical flow, and early and late windows used by the referenced protocol.Its notation identifies frame t in a T-frame sequence, masks Ωstatic and Ωflow, flow ut from It to It+1, and windows E and L.Source: S1. Formal Factor Definitions · p. 11
  • Feature-aligned Background Drift uses matched ORB pairs between the first frame and frame t, and fBD is expressed as a percentage of image size.The image-size normalization uses the frame diagonal d.Source: S1. Formal Factor Definitions · p. 11
  • Normalized Background Flow is normalized by frame width and the interval of sampled pairs, making it independent of native frame rate and sampling policy.The passage identifies W as frame width and ∆t as the interval of pairs actually used for flow estimation.Source: S1. Formal Factor Definitions · p. 11
  • Global-motion compensation fits a robust similarity transform to static-region correspondences and uses its induced displacement field.The transform is written as Tt(x) = stRtx+bt, with displacement dt(x) = Tt(x) −x.Source: S1. Formal Factor Definitions · p. 11
  • The foreground-flow persistence ratio is clipped at a fixed c so a vanishing early window cannot produce an unbounded ratio.S1 also names motion-compensated foreground flow and flow persistence as a factor.Source: S1. Formal Factor Definitions · p. 11
  • Drift leakage and attenuation use late-window mean flow magnitudes over static and flow regions, comparing static flow with raw and compensated flow.DLR is deliberately unbounded, not a fraction; DLR > 1 means static support moves more than the region intended to move. DAR is stored signed and reported clipped to [0, 1].Source: S1. Formal Factor Definitions · p. 11

S2. Region Partition

SNF-Bench partitions each sequence into dynamic and static regions from early optical flow, excluding uncertain boundary pixels from both. The estimate reduces circularity from late drift but remains dependent on early checkpoint behavior and the flow estimator.

Region Partition

  • The partition thresholds mean optical-flow magnitude over the leading 12% of frame pairs with Otsu’s method, then erodes the static region and excludes a 4% frame border.The resulting transition band contributes to neither region; the same procedure is used for both tracks.Source: S2. Region Partition · p. 11
  • Using an early window reduces circularity from accumulated late drift; at a 60 s horizon, the partition is fixed from the first 7.1 s.It remains dependent on early checkpoint behavior and the same flow estimator used downstream.Source: S2. Region Partition · p. 11
  • Factor admission requires signed target-family Spearman correlation ρ ≥0.7 and off-target response below 25%; rotational response is reported but not gated.For rotation, DLR is inverted while fBD and DAR carry that case; VBench DD is a contrast row, not a candidate factor.Source: Table S1 · p. 12
  • Because partitions are automatic rather than annotated, no inter-annotator agreement is reported, and image-conditioned partitions are derived per generated sequence rather than from the shared source image.A system drifting within the first 12% can influence its own partition; measurements are recomputed under systematic boundary erosion and dilation, with responses reported in Table S1.Source: S2. Region Partition · p. 12, Table S1 · p. 12

S3. Validation Protocol · S4. Partition and Composition Robustness

The validation protocol makes geometric corruptions comparable by their induced pixel displacement. Partition robustness checks find stable normalized-background-flow ordering without precipitation, but leave possible mask circularity unresolved.

S3. Validation Protocol

  • Geometric severity is measured as mean pixel displacement by rollout end, with translation d pixels, rotation d/¯r radians, and zoom 1 + d/¯r.Here ¯r is the mean distance from the frame center.Source: S3. Validation Protocol · p. 12
  • At evaluation resolution, one degree of rotation induces roughly 2.7 px mean displacement, so native-parameter comparisons can differ in severity by over an order of magnitude.A corruption weaker than baseline drift can make a factor appear non-monotone.Source: S3. Validation Protocol · p. 12

S4. Partition and Composition Robustness

  • The partition audit asks whether systems benefit from easier masks and whether ordering depends on precipitation, the scene category the partition cannot represent.Source: S4. Partition and Composition Robustness · p. 12
  • Static support occupies a similar 0.61–0.66 share of the frame on average across audited systems, so none is scored over markedly different support areas.Source: S4. Partition and Composition Robustness · p. 12
  • Support area correlates negatively with feature-based drift across seven systems, but the correlation is undetermined and its sign is consistent with possible mask circularity.A system drifting within the estimation window could enlarge its dynamic region and shrink the support over which its drift is measured.Source: S4. Partition and Composition Robustness · p. 12
  • Recomputing scores with a partition not derived from the same sequence is required future work, and the authors do not claim to have run that test.Source: S4. Partition and Composition Robustness · p. 12
  • Removing precipitation leaves normalized-background-flow ordering unchanged and shifts only midfield positions under feature-based drift.Rain and snow cross static support, forcing a two-way partition to assign those pixels by flow magnitude alone.Source: S4. Partition and Composition Robustness · p. 12

S5. Paired Differences · S6. Generation Configuration · S7. Measurement Coverage

The supplement clarifies how paired comparisons, recorded generation settings, and valid-measurement coverage constrain interpretation of the audit. It also reports partition robustness checks and limits, including the distinction between common-configuration comparisons and released-pipeline outputs.

S5. Paired Differences

  • Separation claims use bootstrap distributions of per-prompt paired differences, since overlapping or non-overlapping marginal intervals can misstate the supported comparison.Systems share prompts, making the paired difference the statistic supported by the design.Source: S5. Paired Differences · p. 12

S6. Generation Configuration

  • All text-conditioned checkpoints use a recorded four-step, six-frame-block configuration with seed 0; outputs are 832 × 480 at 16 fps, and guidance scale 5.0 is not applied.Only checkpoint weights and recorded regular/EMA selection vary within a horizon; the conditional-only pipeline does not read the YAML guidance-scale field.Source: S6. Generation Configuration · p. 12
  • The text-conditioned sweep is a reproducible common-configuration comparison, not an audit under each method’s released defaults or operating point.Several checkpoints are distilled for a specific step budget; evaluating all at four steps does not observe them at their own operating point, and rerunning at released defaults is the stated next audit version.Source: S6. Generation Configuration · pp. 12–13
  • The partition derives automatically from early-window flow with an ignored transition band, but its two-way design cannot represent layered content such as rain over static support.This is identified as a scope limit of the partition.Source: Figure S1 · p. 13
  • Across audited systems, static-region area is similar, but its negative association with drift across seven systems is undetermined and flagged as an open validity question.The sign is consistent with possible circularity: early drift could enlarge a system’s dynamic region.Source: Table S2 · p. 13
  • Removing precipitation, which the two-way partition cannot represent, leaves the normalized-background-flow ordering unchanged.Source: Table S2 · p. 13
  • Image-conditioned rows without † are released-pipeline outputs, while † rows use a common rollout wrapper and are treated as deployment sensitivity, not ranked against them.The wrapper uses six frames per block and a fixed seed; denoising schedules remain checkpoint-specific, and wrapper parameters do not apply to unmarked rows.Source: S6. Generation Configuration · p. 13

S7. Measurement Coverage

  • Coverage counts valid per-video measurements so early-terminated runs are partial, while fBD abstentions for too few repeatable keypoints are tabulated rather than hidden in a smaller n [26].fBD is feature-based.Source: S7. Measurement Coverage · p. 13

S8. Complete Audit, Both Tracks · S9. Aggregation Sensitivity · S10. Whole-Frame and SNF-Bench Rank Correlation

The audit reports paired NBF comparisons across systems, tests whether aggregation changes the rankings, and documents method-level rank-correlation analyses for whole-frame and SNF-Bench measures.

S8. Complete Audit, Both Tracks

  • At 60s and 120s, both T2V and I2V tracks are reported; 5s is an initialization check and 240s an exploratory stress horizon.Uncertainty is estimated with a percentile bootstrap over items. I2V dagger-marked wrapper rows are deployment-sensitivity results, not published-method performance.Source: S8. Complete Audit, Both Tracks · p. 13
  • -83.98 was the mean NBF difference for Infinite-Forcing −Causal-Forcing at 60s, and all six paired-difference intervals excluded zero.The other mean differences were Infinite-Forcing −LongLive: -10.22 [−14.72, −6.18]; Infinite-Forcing −Self-Forcing: -9.07 [−13.29, −5.53]; Infinite-Forcing −Rolling-Forcing: -7.58 [−10.38, −5.12]; Infinite-Forcing −CausVid: -5.62 [−9.61, −2.10]; and Infinite-Forcing −Reward-Forcing: -4.86 [−7.48, −2.69].Source: S8. Complete Audit, Both Tracks · p. 14, Table S3 · p. 14
  • Every audited T2V checkpoint uses a recorded common configuration, so these results do not reconstruct the methods’ released inference procedures.The configuration uses a four-step denoising schedule with released-scheduler warping, six frames per block, and seed 0; NBF, fBD, and flow-estimation rate are normalized for comparison.Source: Table S4 · p. 14

S9. Aggregation Sensitivity

  • The best- and worst-ranked systems remain the same under prompt averaging and equal category weighting, but intermediate ranks move.Prompt averaging is the main-paper aggregation; equal category weighting would give two-prompt cells the same influence as a fourteen-prompt cell, increasing variance, so no ordinal claim is made about middle-ranked systems.Source: S9. Aggregation Sensitivity · p. 14

S10. Whole-Frame and SNF-Bench Rank Correlation

  • Tables S11 and S12 report method-level rank correlations and interpretation changes that satisfy the fixed inclusion rule.The supplied passage does not give correlation values or identify which interpretations changed.Source: S10. Whole-Frame and SNF-Bench Rank… · p. 14
  • In I2V, unmarked rows are released-pipeline outputs, while dagger-marked rows use a common wrapper and are not ranked together with them.The wrapper uses six frames per block, a fixed seed, and checkpoint-specific denoising schedules.Source: Table S5 · p. 15
  • fBD is withheld only for a desert dust-storm scene with too few repeatable keypoints; the other factors remain reported for those clips.The flat, dust-obscured ground provides little stable structure to track.Source: Table S6 · p. 15

S11. Deployment Sensitivity · S12. Signed Drift Attenuation · S13. Evaluation-Set Composition

These supplementary sections document deployment-wrapper changes, explain how negative signed drift attenuation values are handled, and describe the evaluation-set coverage reported by category, track, and horizon.

S11. Deployment Sensitivity

  • Table S13 reports changes from released-pipeline outputs to the common long-horizon wrapper for checkpoints evaluated under both settings.The passage gives no values for those changes.Source: S11. Deployment Sensitivity · p. 15

S12. Signed Drift Attenuation

S13. Evaluation-Set Composition

S14. Additional Qualitative Material

The supplementary material compares whole-frame motion with background-flow rankings across horizons and adds qualitative views of T2V and I2V outputs, factor profiles, and scope boundaries. Its audit tables cover four horizons while cautioning against comparisons across differing prompt sets or generation settings.

Complete audits

  • The complete T2V and I2V audits span four horizons, but prompt sets differ across horizons; I2V released-pipeline and wrapper outputs are reported separately rather than ranked against one another.The 5 s tier is an initialization check, while 240 s is a diagnostic extreme for T2V and an exploratory stress horizon for I2V.Source: Table S8 · p. 16, Table S9 · p. 17

Deployment sensitivity

  • Deployment sensitivity reports changes from each checkpoint’s own released setting to the common long-horizon wrapper for the two checkpoints evaluated both ways, without treating wrapper scores as published-method performance.Source: Table S13 · p. 19

Image-conditioned qualitative comparison

  • On a shared source image, I2V systems begin from the identical frame, so rollout divergence is attributable to the system; released-pipeline and wrapper outputs are separated and ordered by fBD only within each group.This attribution does not apply to T2V, where systems author their own scene layouts.Source: Figure S3 · p. 20

Factor profiles

  • No factor profile encloses the others on either track, and the profiles are roster-relative rather than absolute comparisons.Image-conditioned wrapper outlines are shown for profile shape, not as peers of released-pipeline outlines.Source: Figure S4 · p. 21

Scope boundaries

  • The scope illustrations show that a two-way partition cannot separate precipitation from support, near-best static fidelity can coincide with stopped intended motion, and high leakage with low attenuation does not identify the residual motion.Source: Figure S5 · p. 21

Rank disagreement

  • Figure S7 compares whole-frame motion and normalized background-flow rankings at every audited horizon, marking horizons where Dynamic Degree is available only as a method-level aggregate.At 60 s, rank disagreement is described over seven method means and is explicitly descriptive, not inferential; smoothness and flickering share identical ranks.Source: Table S11 · p. 18, Figure S7 · p. 23

Related papers

Loading 2608.28694v1…