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

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

cs.CVAug 27, 2026

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

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

Organizations: Sogang University Seoul, Korea

Abstract

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/.

Figures & tables

Figure 1: Why a fixed-camera benchmark needs a spatial decomposition. Three $60$ s outputs from released checkpoints under the recorded common configuration. A whole-frame motion score rates the first two rows almost identically, yet only the first holds its scene in place; their static-region drift differs by over an order of magnitude. The third row shows the opposite failure, which that score does detect. DD ∗ is the per-video quantity underlying VBench [ 12 ] Dynamic Degree, computed under its published rule but read before the binarization the suite applies when aggregating, so that near-identical rows remain distinguishable as numbers.

Figure 2: Mechanistic validation. Curves are relative to unperturbed fixed-camera clips. Translation raises fBD and NBF; progressive late freezing collapses MCFF-L and FP; under the same translation, VBench Dynamic Degree changes little and slightly upward. The perturbations are synthetic and implicate no published method.

Factor	target	$\rho$	$\rho_{\mathrm{rot}}$	off-target	admitted
fBD	drift	+1.00	+0.70	14%	✓
NBF	drift	+1.00	+0.60	6%	✓
MCFF-L	freeze	+1.00	n/a	14%	✓
FP	freeze	+0.80	n/a	21%	✓
DLR	translational leakage	+1.00	-0.80	5%	✓
DAR	drift	+0.90	+1.00	18%	✓

Table 1: Mechanistic validation and factor admission.

\rho

is signed target-family Spearman correlation;

\rho_{\mathrm{rot}}

reports rotational response; off-target is the largest relative excursion under photometric and mask perturbations. Admission requires

\rho\geq 0.7

and off-target response below

25\%

. Rotation is reported but not gated: DLR is inverted there, while fBD and DAR carry that case. VBench DD is a contrast row, not a candidate factor. Twelve clips per severity.

Check	result
Static-mask area	0.614–0.663
Area–fBD Spearman	$\rho=-0.75$ , exact $p=0.066$ , $n=7$
No precipitation: NBF order	identical
No precipitation: fBD extremes	unchanged
Boundary perturbation: max. response	$\leq 14\%$

Table 2: Partition robustness at 60 s. Static-region area is similar across checkpoints, and the headline NBF ordering is unchanged when precipitation is excluded. The negative area–fBD association has the direction expected under early-mask circularity and is therefore reported as an open validity concern rather than dismissed.

Figure 3: T2V comparison under the recorded common configuration. Rows show t=0,30,60 s for one shared prompt; columns are ordered by increasing measured fBD. Green and orange borders identify the CausVid/Causal-Forcing pair highlighted in Figure 1; the fixed selection rule is documented in Supplementary Section S14.

60s horizon ( $n=23$ )
Method	fBD $\downarrow$	NBF $\downarrow$	MCFF E $\rightarrow$ L	FP $\uparrow$
CausVid	8.08 $\pm$ 3.15	9.27 $\pm$ 4.14	2.87 $\rightarrow$ 1.96	0.67 $\pm$ 0.18
Self-Forcing	11.4 $\pm$ 3.63	12.7 $\pm$ 3.97	10.4 $\rightarrow$ 2.54	0.44 $\pm$ 0.22
Infinite-Forcing	7.75 $\pm$ 3.84	3.65 $\pm$ 0.88	1.17 $\rightarrow$ 0.29	0.63 $\pm$ 0.22
Rolling-Forcing	16.8 $\pm$ 3.69	11.2 $\pm$ 2.70	3.18 $\rightarrow$ 1.57	0.75 $\pm$ 0.29
Reward-Forcing	14.8 $\pm$ 3.71	8.51 $\pm$ 2.79	3.81 $\rightarrow$ 0.80	0.49 $\pm$ 0.18

Table 3: Recorded common-configuration T2V audit. A four-step denoising schedule with released-scheduler warping, six frames/block, and seed

0

. Header

n

is the maximum valid item count within each horizon; factor-specific abstentions are in Table S6. Units: fBD is % frame diagonal; NBF is

10^{-3}

frame widths/s; MCFF is px/sampled interval; FP is unitless. FP is the mean of per-video clipped late/early ratios and therefore need not equal the ratio of the displayed aggregate MCFF-L and MCFF-E means. Means with percentile-bootstrap 95% CI half-width shown for compactness (10k resamples, seed 0);

n<10

shows means only. MCFF-E

\rightarrow

L accompanies FP. Separation uses paired per-prompt intervals (Table S3). DLR/DAR are supplementary.

60s horizon ( $n=30$ )
Method	fBD $\downarrow$	NBF $\downarrow$	MCFF E $\rightarrow$ L	FP $\uparrow$
Causal-Forcing++ (2-step) †	9.18	5.97	0.85 $\rightarrow$ 0.60	0.86
Causal-Forcing++ (1-step) †	5.34	7.05	0.86 $\rightarrow$ 0.57	0.77
Causal-Forcing (frame-wise) †	21.1	36.4	2.80 $\rightarrow$ 2.69	0.89
Causal-Forcing++ (2-step, frame-wise)	9.94	6.91	1.30 $\rightarrow$ 0.78	1.02
Self-Forcing †	17.3	5.52	1.50 $\rightarrow$ 0.69	0.89

Table 4: I2V released pipelines and deployment sensitivity at 60 and 120 s. Unmarked rows are released-pipeline outputs; † rows use the wrapper; settings are not compared. Header

n

is the maximum valid item count within each horizon; factor-specific abstentions are in Table S6. Units: fBD is % frame diagonal; NBF is

10^{-3}

frame widths/s; MCFF is px/sampled interval; FP is dimensionless. FP is the mean of per-video clipped late/early ratios and therefore need not equal the ratio of the displayed aggregate MCFF-L and MCFF-E means. Means are shown; DLR and DAR are reported in Supplementary Table S9. MCFF-E→L accompanies FP, and released-pipeline and wrapper rows are not compared.

Figure 4: I2V comparison (60s) on a shared source (0s). Unmarked outputs are produced by released pipelines; † marks outputs generated with the common wrapper.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1 : Region-partition protocol. (a) The partition is derived automatically from early-window flow, with an ignored transition band at the boundary. (b) A two-way partition cannot represent layered content such as rain crossing static support, which is the scope limit stated in the main paper.

Factor	target	$\rho$	$\rho_{\mathrm{rot}}$	off-target	admitted
fBD	drift	+1.00	+0.70	14%	✓
NBF	drift	+1.00	+0.60	6%	✓
MCFF-L	freeze	+1.00	n/a	14%	✓
FP	freeze	+0.80	n/a	21%	✓
DLR	translational leakage	+1.00	-0.80	5%	✓
DAR	drift	+0.90	+1.00	18%	✓

Appendix

Table S1 : Mechanistic validation and factor admission.

\rho

is signed target-family Spearman correlation;

\rho_{\mathrm{rot}}

reports rotational response; off-target is the largest relative excursion under photometric and mask perturbations. Admission requires

\rho\geq 0.7

and off-target response below

25\%

. Rotation is reported but not gated: DLR is inverted there, while fBD and DAR carry that case. VBench DD is a contrast row, not a candidate factor. Twelve clips per severity.

System	static-region area
CausVid	0.663 $\pm$ 0.073
Reward-Forcing	0.661 $\pm$ 0.060
Self-Forcing	0.651 $\pm$ 0.087
Infinite-Forcing	0.650 $\pm$ 0.090
LongLive	0.644 $\pm$ 0.095
Rolling-Forcing	0.616 $\pm$ 0.107

Appendix

Table S2: Partition and composition robustness. The static region occupies a similar share of the frame for every audited system, so no system is scored over a much larger or smaller support than another. The area does correlate negatively with drift across systems; over seven systems that correlation is not determined, but its sign is the one circularity would produce—a system that drifts early enlarges its own dynamic region—and we flag it as an open validity question rather than dismissing it. Removing precipitation, the category a two-way partition cannot represent, leaves the normalized-background-flow ordering unchanged.

Comparison	$n$	mean diff.	95% CI
Infinite-Forcing $-$ Causal-Forcing	23	-83.98	$[-102.59,\,-65.52]$
Infinite-Forcing $-$ LongLive	23	-10.22	$[-14.72,\,-6.18]$
Infinite-Forcing $-$ Self-Forcing	23	-9.07	$[-13.29,\,-5.53]$
Infinite-Forcing $-$ Rolling-Forcing	23	-7.58	$[-10.38,\,-5.12]$
Infinite-Forcing $-$ CausVid	23	-5.62	$[-9.61,\,-2.10]$
Infinite-Forcing $-$ Reward-Forcing	23	-4.86	$[-7.48,\,-2.69]$

Appendix

Table S3: Paired differences in NBF, 60s horizon. Every system is evaluated on the same prompts, so separation is judged on the bootstrap distribution of per-prompt differences (10k resamples, fixed seed) rather than on whether two marginal intervals happen to overlap—overlapping intervals can hide a reliable paired difference, and non-overlapping ones can suggest a difference that is not. 6 of 6 intervals exclude zero.

System	setting	resolution	fps	horizons
CausVid	common	832 $\times$ 480	16	5s, 60s, 120s, 240s
Self-Forcing	common	832 $\times$ 480	16	5s, 60s, 120s, 240s
Infinite-Forcing	common	832 $\times$ 480	16	5s, 60s, 120s, 240s
Rolling-Forcing	common	832 $\times$ 480	16	5s, 60s, 120s, 240s
Reward-Forcing	common	832 $\times$ 480	16	5s, 60s, 120s, 240s
LongLive	common	832 $\times$ 480	16	5s, 60s, 120s, 240s

Appendix

Table S4: Recorded generation configuration (T2V). Every audited system was run from its released checkpoint. Output formats are normalized at measurement time: NBF is expressed per second and per frame width, fBD as a percentage of the frame diagonal, and every sequence is resampled to a common flow-estimation rate. All rows use the recorded common T2V configuration: a four-step denoising schedule with released-scheduler warping, six frames per block, and seed

0

. These are checkpoint results under this common configuration, not reconstructions of the methods’ released inference procedures.

System	setting	resolution	fps	horizons
Causal-Forcing++ (2-step, frame-wise)	released	832 $\times$ 480	16	5s, 60s, 120s, 240s
Wan2.1-I2V-14B-480P	released	832 $\times$ 464	16	5s
Wan2.2-I2V-A14B	released	832 $\times$ 464	16	5s
LTX-Video 13B-0.9.8-distilled	released	832 $\times$ 480	24	5s
Causal-Forcing++ (1-step, frame-wise)	released	832 $\times$ 480	16	60s, 120s
Causal-Forcing++ (2-step) †	wrapper	832 $\times$ 480	16	5s, 60s, 120s, 240s

Appendix

Table S5: Recorded generation configuration (I2V). Every audited system was run from its released checkpoint. Output formats are normalized at measurement time: NBF is expressed per second and per frame width, fBD as a percentage of the frame diagonal, and every sequence is resampled to a common flow-estimation rate. Unmarked rows are recorded outputs from model-specific released pipelines and carry the setting label released . † rows use the common long-horizon rollout wrapper (six frames per block, fixed seed; checkpoint-specific denoising schedules); those parameters do not apply to unmarked rows, and the two settings are not ranked together.

track	horizon	measured	fBD abstained	rate
I2V	120s	140	0	0.0%
I2V	240s	30	0	0.0%
I2V	5s	90	0	0.0%
I2V	60s	210	2	1.0%
T2V	120s	42	0	0.0%
T2V	240s	28	0	0.0%

Appendix

Table S6: Public-roster measurement coverage and fBD abstention. fBD is feature-based and declines to report when too few repeatable keypoints survive in the static region. Public-roster abstentions are confined to a desert dust-storm scene whose flat, dust-obscured ground offers little stable structure to track. The remaining factors are reported for those clips as usual; only fBD is withheld.

Method	status	setting	5s	60s	120s	240s
CausVid	public	common	V6 T6 B	V23 T23 B	V6 T6 B	V4 T4 B
Self-Forcing	public	common	V6 T6 B	V23 T23 B	V6 T6 B	V4 T4 B
Infinite-Forcing	public	common	V6 T6 B	V23 T23 B	V6 T6 B	V4 T4 B
Rolling-Forcing	public	common	V6 T6 B	V24 T24 B	V6 T6 B	V4 T4 B
Reward-Forcing	public	common	V12 T12 B	V23 T23 B	V6 T6 B	V4 T4 B
LongLive	public	common	V6 T6 B	V23 T23 B	V6 T6 B	V4 T4 B

Appendix

Table S7: Asset and score coverage for the audited public systems. V

n

is the number of available video assets, T

n

the number of valid SNF task-metric records, and B denotes available VBench scores. T

n

N

marks entries where only

n

N

videos hold usable measurements. V counts generated assets, whereas audit

n

counts unique frozen-manifest prompt items after duplicate prompt outputs are collapsed. Reward-Forcing has 12 assets over six 5 s prompts; Rolling-Forcing has 24 assets over 23 60 s prompts.

5s horizon ( $n=6$ )
Method	fBD $\downarrow$	NBF $\downarrow$	MCFF-E	MCFF-L $\uparrow$	FP $\uparrow$	DLR $\downarrow$	DAR $\downarrow$
CausVid	2.08	6.56	1.44	2.91	1.24	0.26	0.07
Self-Forcing	0.68	10.8	7.70	4.35	0.76	0.49	0.03
Infinite-Forcing	0.24	2.32	1.23	0.89	1.08	0.21	0.09
Rolling-Forcing	3.51	12.1	3.10	3.69	1.32	0.48	0.17
Reward-Forcing	1.10	5.21	0.63	0.58	1.37	0.44	0.28

Appendix

Table S8: Complete T2V audit, all four horizons. The main paper reports the 60 s and 120 s panels; the 5 s tier is an initialization check and 240 s a diagnostic extreme. Every row is computed under the same frozen specification, so panels are comparable within a horizon; prompt sets differ across horizons, so columns are not comparable between panels. Header

n

is the maximum valid item count within each horizon; factor-specific abstentions are in Table S6. Units: fBD is % frame diagonal; NBF is

10^{-3}

5s horizon ( $n=10$ )
Method	setting	fBD $\downarrow$	NBF $\downarrow$	MCFF-E	MCFF-L $\uparrow$	FP $\uparrow$	DLR $\downarrow$	DAR $\downarrow$
Causal-Forcing++ (2-step)	wrapper	3.69	10.0	1.06	0.61	0.78	0.67	0.32
Causal-Forcing++ (1-step)	wrapper	5.28	11.4	1.05	0.47	0.64	0.66	0.29
Causal-Forcing (frame-wise)	wrapper	16.1	45.3	8.48	2.79	0.43	0.68	0.29
Causal-Forcing++ (2-step, frame-wise)	released	1.22	10.6	2.23	0.78	0.47	0.68	0.20
Self-Forcing	wrapper	9.66	5.74	0.69	0.26	0.70	0.74	0.21

Appendix

Table S9: Complete I2V audit, all four horizons. The main paper reports the 60s and 120s panels; the 5s tier is an initialization check and the 240s tier is an exploratory stress horizon. All rows are scored with the same frozen measurement specification; however, released-pipeline and wrapper rows represent different generation settings and are reported separately rather than ranked against one another. Prompt sets differ across horizons, so values are not compared across panels. Header

n

is the maximum valid item count within each horizon; factor-specific abstentions are in Table S6. Units: fBD is % frame diagonal; NBF is

10^{-3}

Figure S2 : Operating regimes under the recorded common T2V configuration. Each checkpoint in the static-drift–surviving-flow plane. Color encodes low-to-high attenuation of measured dynamic-region flow after global-motion compensation, not a share attributable to drift: the quantity is signed, and values are shown clipped to

[0,1]

. The view exposes trade-offs; it is not a leaderboard.

System	prompt mean	rank	macro mean	rank
NBF
Infinite-Forcing	3.65	1	4.12	1
Reward-Forcing	8.51	2	10.11	4 †
CausVid	9.27	3	9.93	3
Rolling-Forcing	11.23	4	11.32	5 †
Self-Forcing	12.72	5	8.98	2 †

Appendix

Table S10: Sensitivity of the audit to category aggregation. Prompt means, as reported in the main paper, against category macro-averages, for the two static-fidelity factors at the principal horizon. † marks a system whose rank changes. The best- and worst-ranked systems are identical under both, so the paper’s headline contrast does not depend on the choice; intermediate positions do change, which is why no ordinal claim is made about the middle of the field.

generic metric	fBD	NBF	MCFF	DLR
VB-DD	+0.75	+0.93	+0.89	+0.36
VB-bg	-0.64	-0.71	-0.75	-0.50
VB-smooth / VB-flick	-0.96	-0.82	-0.61	-0.64

Appendix

Table S11: Rank disagreement at 60 s. Spearman correlation over

n=7

method means—descriptive, not inferential. Positive correlation between apparent motion and background flow is the effect the benchmark isolates. Smoothness and flickering rank these systems identically and share a row.

Method	Generic evidence	SNF-Bench evidence	Interpretation
CausVid	VB-DD rank 6/7	NBF rank 3/7, MCFF-L rank 4/7	stable sequence with decaying intended motion
Causal-Forcing	VB-DD rank 1/7	NBF rank 7/7, MCFF-L rank 1/7	high apparent motion accompanied by high static-region motion under the recorded common configuration
Infinite-Forcing	VB-DD rank 7/7	NBF rank 1/7, MCFF-L rank 7/7	stable sequence with decaying intended motion
LongLive	VB-DD rank 3/7	NBF rank 6/7, MCFF-L rank 3/7	high apparent motion accompanied by high static-region motion under the recorded common configuration
Reward-Forcing	VB-DD rank 5/7	NBF rank 2/7, MCFF-L rank 6/7	stable sequence with decaying intended motion
Self-Forcing	VB-DD rank 2/7	NBF rank 5/7, MCFF-L rank 2/7	high apparent motion accompanied by high static-region motion under the recorded common configuration

Appendix

Table S12: Interpretation changes between generic whole-frame metrics and SNF-Bench at 60 s. Neither evaluation is labeled correct; the table identifies information hidden by whole-frame aggregation. Cases are selected by a fixed rank-gap rule applied to all public methods. Rolling-Forcing does not satisfy the fixed three-rank-gap inclusion criterion and is therefore absent.

Checkpoint	horizon	$\Delta$ fBD	$\Delta$ NBF	$\Delta$ MCFF-L	$\Delta$ DLR
Causal-Forcing++ (2-step)	5s	+2.47	-0.61	-0.17	-0.01
Causal-Forcing++ (2-step)	60s	-0.76	-0.95	-0.17	-0.08
Causal-Forcing++ (2-step)	120s	-1.71	-8.46	-0.73	-0.17
Causal-Forcing++ (2-step)	240s	+4.06	+0.88	-0.09	-0.28
Causal-Forcing++ (1-step)	60s	-5.05	+1.80	+0.05	-0.20
Causal-Forcing++ (1-step)	120s	-6.34	+2.00	+0.31	-0.14

Appendix

Table S13: Deployment sensitivity. Change from each checkpoint’s own released setting to the common long-horizon wrapper, for the two checkpoints evaluated both ways. Absolute wrapper scores are never read as the published method’s performance.

Method	track	dur	n	negative	rate	most negative
CausVid	i2v	120s	20	1	5.0%	-0.055
CausVid	i2v	5s	10	5	50.0%	-1.076
CausVid	i2v	60s	30	2	6.7%	-0.084
Causal-Forcing++ (1-step)	i2v	120s	20	1	5.0%	-0.027
Causal-Forcing (frame-wise)	i2v	120s	20	3	15.0%	-0.171
Causal-Forcing++ (2-step)	i2v	5s	10	1	10.0%	-0.164

Appendix

Table S14: Incidence of negative DAR, i.e. clips where global-motion compensation increased measured dynamic-region flow. 21 further entries had no negative values and are omitted.

track/duration	n	river_stream	ocean_waves	precipitation	fire_smoke	lava_volcanic	windborne	cats
i2v/5s	10	2	·	3	2	1	2	5
i2v/60s	30	9	4	9	1	3	4	6
i2v/120s	20	6	3	3	3	1	4	6
i2v/240s	5	1	2	1	·	1	·	4
t2v/5s	6	3	1	1	1	·	·	4
t2v/60s	23	14	2	3	2	·	2	5

Appendix

Table S15 : Scene-category balance of the evaluation set, by track and horizon. A dot marks a category absent from that cell.

Figure S3 : Image-conditioned systems on one shared source image. Every system begins from the identical frame, so divergence over the rollout is attributable to the system rather than to a differently imagined scene—which is not true of the text-conditioned track, where each system authors its own layout. Released-pipeline and wrapper outputs are visibly separated and ordered by increasing fBD only within their respective groups; † marks a wrapper output, and no cross-setting ordering is intended.

Figure S4 : Factor profile per track. Each factor is rescaled across the audited systems and oriented so that outward is always better; plotting raw values would place “most drift” and “most surviving motion” both outward and the shape would carry no meaning. No system encloses the others on either track, which is the conclusion the operating-regime view reaches, shown as a profile. On the image-conditioned panel, dashed outlines marked † are rows run under the common long-horizon wrapper rather than a released pipeline: they are shown for shape, and are not peers of the solid released-pipeline outlines. The rescaling is relative to the audited roster, so a profile is descriptive of this set and would change if the set changed.

Figure S5 : Stated scope boundaries. (a) A two-way partition cannot separate precipitation from the support it crosses. (b) A system can attain near-best static fidelity precisely because its intended motion has also stopped. (c) High leakage with low attenuation marks static-region motion for which the fitted global similarity model attenuates little measured flow; what the residual consists of is not identified by these measurements.

Figure S6 : T2V comparison under the recorded common configuration. Every audited text-conditioned checkpoint is evaluated on the same prompt. Rows show t=0,30,60 s; columns are ordered by increasing measured fBD.

Figure S7 : Rank disagreement at every audited horizon. Rank under a whole-frame motion score against rank under normalized background flow. Both quantities are independent of the global-motion estimator; prompt counts are given per panel. An asterisk marks a horizon where Dynamic Degree is available only as a method-level VBench aggregate, suitable for descriptive ranks but not for item-level uncertainty.

Figure S8 : Composition of the evaluation set by scene category, in prompts per category at each track and horizon.

Video Generation Evaluation

Video generation evaluation measures visual quality, prompt adherence, temporal consistency, physical plausibility, and task-specific behavior in synthesized videos.

See all 192 papers

Steady-Forcing: Balancing Spatial Persistence and Motion Continuity in Long-Horizon Nature Video Diffusion
Jun 2, 2026Matiur Rahman Minar, Seunghun Oh, GangHyeon Jeong +1Physical Consistency in Video Generation Long-Horizon Video Generation
Department of Computer Science and Engineering Sogang University Seoul, Korea · Department of Artifical Intelligence Sogang University Seoul, Korea
Towards Subject Consistency over Dynamic Subject Sets in Video Generation
Oct 1, 2026Tongcheng Zhang, Jun Zhu, Jianfei ChenLong Video Generation Identity-Preserving Video Generation
Dept. of Comp. Sci. and Tech., Institute for AI, BNRist Center, THBI Lab, Tsinghua-Bosch Joint ML Center, Tsinghua University
Goodbye Drift: Anchored Tree Sampling for Long-Horizon Video-to-Video Generation
May 19, 2026Matthew Bendel, Stephen W. Bailey, Mithilesh Vaidya +2Video-Reference Video Generation Long-Horizon Video Generation
Descript, Inc.

Jun 2, 2026J/K move · Enter open · S save

Steady-Forcing: Balancing Spatial Persistence and Motion Continuity in Long-Horizon Nature Video Diffusion

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

Department of Computer Science and Engineering Sogang University Seoul, Korea · Department of Artifical Intelligence Sogang University Seoul, Korea

Physical Consistency in Video Generation Long-Horizon Video Generation Memory-Augmented Video Generation Video Generation Autoregressive Video Generation Diffusion Models for Motion Generation+1

Open paper AskPDF arXiv

Autoregressive video diffusion models enable streaming generation but often degrade over long rollouts: static scene layouts drift, while mechanisms that improve spatial stability tend to suppress motion, causing natural flows such as water, fire, or smoke to stagnate. We study this stability-motion trade-off in fixed-camera long-horizon nature video generation, where the two failure modes can be more clearly separated than in moving-camera settings. We propose Steady-Forcing, a memory and training framework combining a persistent visual anchor (V-Sink), an exponential moving-average motion memory (EMA-Sink), block-relative temporal encoding, periodic cache purification, and distillation from a Wan2.1-14B teacher with motion-rewarded priors under task-focused configurations. Together, these components are designed to preserve background identity while sustaining visually plausible fluid dynamics over multi-minute autoregressive rollouts. Evaluations across seven baselines show that Steady-Forcing improves long horizon background consistency and imaging quality, while a blind user study indicates stronger perceived stability and motion continuity. The benchmark evaluation further suggest that generic VBench aggregate scores under-penalize fixed-camera artifacts as well as rewarding drift-induced optical flow as Dynamic Degree while not directly penalizing texture hardening or flow stagnation - motivating future task-specific benchmarks for static-camera nature-flow evaluation. Project page: https://minar09.github.io/steadyforcing/

Abstract

Video Generation Evaluation

Explore similar work

Steady-Forcing: Balancing Spatial Persistence and Motion Continuity in Long-Horizon Nature Video Diffusion

Towards Subject Consistency over Dynamic Subject Sets in Video Generation

Goodbye Drift: Anchored Tree Sampling for Long-Horizon Video-to-Video Generation