Pith. sign in

REVIEW 3 major objections 5 minor 67 references

Autoregressive video models can be steered away from dead-end trajectories at test time by checking whether their predicted noise still looks like Gaussian noise.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 22:08 UTC pith:33L4DMMH

load-bearing objection Useful empirical test-time recipe; the terminal-point detection story is not backed by the evidence and the stress-test concern is real. the 3 major comments →

arxiv 2607.15849 v1 pith:33L4DMMH submitted 2026-07-17 cs.CV cs.AI

Test-Time Noise Guided Adaptation for Realistic Autoregressive Video Generation

Dimitrios Karageorgiou , Symeon Papadopoulos , Ioannis Kompatsiaris , Efstratios Gavves This is my paper
classification cs.CV cs.AI
keywords autoregressive video generationterminal pointstest-time adaptationnoise consistency hypothesisdiffusion modelsflow matchinglow-rank adaptationvideo realism
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper sets out to explain a failure mode of autoregressive video diffusion that previous fixes miss: even when every generated frame individually looks realistic, the joint trajectory can reach a point the model has no learned continuation for, called a terminal point. The authors claim that this state is detectable at test time because the denoiser's one-step-ahead noise prediction stops matching isotropic Gaussian noise. Their method, TANGO, uses that mismatch as a critic signal and then optimizes a lightweight low-rank-adapted copy of the denoiser, constrained near the original prediction, to find an alternative trajectory that avoids the terminal point. The reported payoff is a 3.1% absolute gain in VBench total score over prior autoregressive methods and a 28.3% average reduction in Fréchet Video Distance on 15-second videos, without retraining or extra data. If correct, this offers a general, training-free way to stabilize long video generation inside the model's learned manifold.

Core claim

The central claim is that a terminal point is a trajectory state that lies inside the learned manifold of real frames but has no continuation inside it, and that such states can be detected through the statistics of the predicted noise residual. Under an autoregressive factorization, conditioning on an out-of-distribution past sequence leaves the denoiser free to violate the isotropic Gaussian noise assumption used in training; hence a one-step look-ahead residual that deviates in mean, variance, skewness, kurtosis, spectral flatness, or low-frequency moments signals a terminal point. TANGO turns this into a test-time objective: a low-rank-adapted copy of the denoiser is optimized to produce

What carries the argument

The key mechanism is the noise consistency hypothesis: for well-modeled conditioning sequences, the predicted noise of a diffusion or flow-matching denoiser should match the isotropic Gaussian noise of the forward noising process. TANGO operationalizes this by predicting one step ahead and computing a noise consistency objective over the implied noise residual, penalizing deviations in first and second moments, skewness and excess kurtosis (full-band and low-frequency), and spectral flatness. The diffusion model thereby acts as a frozen critic of its own outputs, while a trainable low-rank (LoRA) copy of the denoiser is optimized under a proximity constraint so the frame stays close to the o

Load-bearing premise

The load-bearing premise is the noise consistency hypothesis: for a well-modeled conditioning sequence the denoiser's predicted noise must look like isotropic Gaussian noise, so any statistically meaningful deviation in a single look-ahead residual flags a terminal point.

What would settle it

Run the same noise-consistency losses on look-ahead residuals obtained when conditioning the model on real, held-out video trajectories: if those residuals violate isotropy as often or as strongly as residuals from generated trajectories, the statistic does not separate terminal points from healthy ones, and the central claim collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Long-video generation can be stabilized without changing the training procedure or adding data, purely by spending inference-time compute.
  • The same noise-consistency check applies to text-, image-, and video-to-video generation, since it does not depend on the conditioning modality.
  • The approach is complementary to scaling: larger models enlarge the space of valid continuations, while noise guidance finds a valid continuation within a fixed learned manifold.
  • Because the adaptive parameters are a low-rank copy updated at test time, the method can be layered onto existing pretrained autoregressive video models with modest memory overhead.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • An extension the paper leaves implicit is calibration: measuring the same noise-consistency statistic on real training trajectories would establish a null distribution and show whether deviations are truly specific to terminal points.
  • The hand-weighted combination of moments and spectral flatness could plausibly be replaced by a single calibrated scalar statistic, making the test-time objective simpler and more sensitive.
  • The same noise-residual test might serve as an early-warning signal for epistemic uncertainty in other autoregressive generative models beyond video, such as world models or long-horizon image generation.
  • If terminal points are a finite-training phenomenon, the benefit of TANGO should shrink as training data and model size grow; measuring the improvement as a function of scale would test the paper's interpretation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper studies error accumulation in autoregressive video diffusion models and attributes it to "terminal points": trajectories whose individual frames lie on the learned manifold but for which no continuation exists inside that manifold. The proposed method, TANGO, is a test-time procedure that adapts a low-rank (LoRA) copy of the denoiser by minimizing a noise-consistency objective computed from a one-step-ahead predicted residual. The objective penalizes deviations from isotropic Gaussian statistics in the mean, variance, higher-order moments, spectral flatness, and low-frequency moments. The adapted prediction is the final output at each autoregressive step. Experiments compare TANGO with five same-scale autoregressive baselines on VBench and LV-Bench FVD, include extensive ablations, runtime/memory analysis, and qualitative examples. The headline quantitative claim is a 3.1% absolute VBench improvement and a 28.3% average FVD reduction on 15-second videos.

Significance. If the terminal-point hypothesis were established, TANGO would be a valuable training-free-at-training stabilization technique that is model-agnostic and applicable to text-, image-, and video-to-video settings. The empirical evaluation is thorough: it uses a consistent base model across all compared methods, reports ablations for every loss term and key hyperparameters, includes runtime and memory comparisons, and the authors commit to releasing code. The ablation results are internally consistent and support the usefulness of each loss component. However, the core claim that the noise-consistency objective specifically detects and avoids terminal points is currently asserted rather than validated. The empirical gains could equally be explained by TANGO acting as a generic test-time regularizer. Because the interpretation and novelty of the paper depend on this hypothesis, the central issue needs substantial additional support before the claims can be accepted as stated.

major comments (3)
  1. [Sec. 3.2 / Eq. (A6)] The noise-consistency hypothesis is not derived and is in tension with the actual training objective. For the flow-matching model used in the experiments, the network is trained to regress the conditional expectation E[epsilon - x0 | x_t, cond], not to sample from the forward noise distribution. Substituting the ideal regression target into Eq. (A6), the predicted residual is r-hat = x_t + (1-t) E[epsilon - x0 | x_t, cond], which equals epsilon only when x0 is deterministic given the conditioning. For stochastic continuations, the residual is a shrunken Gaussian (e.g., variance b^2/(a^2 sigma^2 + b^2) < 1 in a linear-Gaussian model), so the mean/variance term L_a in Eq. (5) penalizes valid in-distribution frames. No calibrated null distribution on real training trajectories is provided to show that L_N separates terminal from non-terminal points. The method may still be a useful regulari
  2. [Sec. 4.1 / Sec. A2] The paper states that hyperparameters (lambda values, LoRA rank, B, learning rate) were tuned 'on the test set of MSR-VTT'. Tuning on a test set is not 'avoiding overfitting'; it is a direct form of test-set leakage. Even if MSR-VTT is not part of the reported evaluation, this procedure invalidates the claim that the hyperparameters are general-purpose and undermines the hyperparameter ablations in Table 3. Use a held-out validation split, or report the sensitivity of the main VBench/FVD results to the chosen hyperparameters. As written, the methodology does not meet standard benchmark practice.
  3. [Sec. 3.3 / Eq. (10)] The adaptation is self-referential: the critic epsilon_theta is the same model being adapted, and the signal used to define a terminal point (the one-step-ahead residual) is produced by that same model. Since 'terminal point' is defined via the model's learned manifold, minimizing L_N could simply move the model to a region where its own residual is more Gaussian without actually avoiding trajectories that have no valid continuation. The FVD/VBench improvements are compatible with both interpretations. A direct test is needed - for example, measuring whether the adapted model reduces the rate at which continuations leave the manifold, or comparing the adapted and unadapted models on sequences deliberately seeded to be terminal. Without such evidence, the causal story remains an unvalidated hypothesis.
minor comments (5)
  1. [Abstract / Sec. 4.3] The 'state-of-the-art' claim is unqualified. The 3.1% VBench gain is relative to same-scale autoregressive methods; Table 4 shows a 27B model (SVI-2.0-Pro) scoring 0.869 versus TANGO's 0.863. Qualify the claim to 'autoregressive baselines of similar scale' or 'test-time methods'.
  2. [Sec. 3.1] The definition of a terminal point uses supp(pD(V)) and the phrase 'no continuation exists inside it' without formalizing how the learned manifold relates to the finite training support. The notion is intuitive but needs a precise definition, especially because the model's learned manifold is continuous and may contain arbitrarily close continuations with low but nonzero probability.
  3. [Eq. (7)] The spectral flatness measure is computed excluding the DC component and with a small epsilon added to the power spectrum. This is a reasonable numerical choice, but it deviates from the standard SFM definition; please justify the choice and report sensitivity to zeta.
  4. [Fig. 1 / Fig. 2] The schematic diagrams are dense and the connection between 'terminal point', 'noise residual', and the optimization loops is not immediately clear. Adding a legend and explicitly marking the critic and adapted branches would improve readability.
  5. [Table 2] The row 'w/o L_N (baseline)' serves as both the no-objective ablation and the baseline for the 28.3% reduction claim. Clarify in the caption that this row is the unadapted SelfForcing model, so readers do not confuse it with a separate baseline method.

Circularity Check

0 steps flagged

No significant circularity: the noise-consistency hypothesis is an explicitly labeled empirical proxy, the terminal-point definition is independent of it, and the headline claims are evaluated on external benchmarks.

full rationale

No circular step meets the paper's own equations or self-citation chain. The terminal point is defined in Sec. 3.1 purely in distributional terms (no continuation exists inside the learned support), and the noise-consistency check in Sec. 3.2 is introduced explicitly as a hypothesis: 'We posit that at a denoising timestep t it is possible to examine whether a predicted x̂0_i constitutes a terminal point by examining whether the predicted look-ahead residuals r̂_{i+1} ... violate the properties of isotropic Gaussian noise.' That is a proposed test-time proxy, not a restatement of the definition. The optimization in Eq. (10) minimizes the noise-consistency objective, but the reported evidence is external: VBench and FVD on LV-Bench, not the optimized loss. Hyperparameters are tuned on MSR-VTT and evaluated on LV-Bench, so no fitted value is renamed as a prediction. Reusing the same diffusion model as generator and critic is a self-referential mechanism, but it is a design choice rather than a definitional reduction, and the paper's central empirical claim is not forced by it. The noise-consistency hypothesis may be theoretically questionable (an MMSE denoiser outputs a conditional expectation, not a noise sample), but that is a correctness/validity risk, not circularity. No load-bearing self-citations by the authors appear; the baseline [22] and related methods are by other groups. Accordingly, the derivation is self-contained enough that no circularity score above 0 is warranted.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 1 invented entities

The central method rests on the unproven noise-consistency hypothesis, a set of hyperparameters tuned on the MSR-VTT test set, and the invented conceptual entity 'terminal point' with no direct independent evidence. The flow-matching algebra itself is standard, but the diagnostic link between Gaussian-looking residuals and trajectory safety is the main unverified load-bearing assumption.

free parameters (3)
  • Noise consistency weights = 0.1, 0.1, 0.1, 0.1, 0.1, 0.2, 0.2, 0.9
    Chosen by preliminary tuning on the MSR-VTT test set under a video-to-video 5s setup; these directly weight the central objective and the regularization.
  • LoRA rank = 8
    Selected from an ablation (2,4,8,16); performance is best at 8 and affects the capacity of the test-time adaptation.
  • Noise vectors B, learning rate, batch size, epochs = B=10, lr=2e-6, batch 2, 1 epoch
    Selected as a compute/quality trade-off; B is ablated and higher B improves FVD further at greater cost.
axioms (5)
  • ad hoc to paper For well-modeled conditioning sequences, the predicted noise of the diffusion model should be isotropic Gaussian (noise consistency hypothesis).
    Introduced in Sec. 3.2 and used to define the objective in Eq. 9; not derived from first principles.
  • ad hoc to paper Deviations of a single predicted residual's empirical moments and spectral flatness from N(0,I) values indicate terminal points.
    Underpins Eqs. 5-8; no null distribution or calibration against residuals from real training trajectories is provided.
  • domain assumption A finite training set implies every sufficiently long generated trajectory must eventually hit a terminal point.
    Sec. 3.1 relies on the support of the empirical training distribution being finite/bounded; this is a conceptual argument, not a theorem.
  • domain assumption Constrained optimization with regularization stays near the original prediction and does not collapse to trivial solutions.
    Sec. 3.3 Eq. 10; the ablation without regularization confirms collapse, but the specific anchor distance is chosen by tuning.
  • standard math Flow matching path and implied noise equations are valid.
    Supp. A1 follows the standard optimal-transport flow matching derivation and is algebraically correct.
invented entities (1)
  • Terminal point no independent evidence
    purpose: Explains collapse in autoregressive generation: a point in the learned real manifold from which no continuation exists inside the manifold; motivates the test-time objective.
    No direct measurement or falsifiable prediction is provided; existence is inferred from downstream metric changes and a conceptual support-set argument.

pith-pipeline@v1.3.0-alltime-deepseek · 3555 in / 6033 out tokens · 147547 ms · 2026-08-01T22:08:01.507740+00:00 · methodology

0 comments
read the original abstract

Autoregressive video diffusion models have enabled the generation of arbitrarily long videos by removing conditioning on future frames, thus greatly improving computational efficiency. Yet, they suffer from error accumulation over time, as the denoised sequence gradually drifts away from the conditioning distribution seen during training. Recent advances attempt to reduce this error by anchoring each generated frame to the learned manifold of real ones. However, even when all generated individual frames lie close to the real manifold, there are trajectories which the model lacks sufficient knowledge to continue without exiting it, thus reaching a terminal point. To prevent the model from being trapped in terminal points, we start from the hypothesis that for well-modeled future trajectories the distribution of the predicted noise should match the one of the forward noising process. To enforce such a prior at test time, we introduce Terminal points Avoidance through Noise Guided Optimization (TANGO), which uses the diffusion model as a critic of its own outputs, by predicting one step forward and requiring an isotropic Gaussian noise prediction. We use the deviation from this expected noise distribution to search for an alternative trajectory that does not lead to a terminal point. Our approach achieves a $3.1\%$ absolute improvement on VBench over state-of-the-art, while reducing Fr\'echet Video Distance by $28.3\%$ on average across $15$s videos. Our code is available on https://mever-team.github.io/tango.

Figures

Figures reproduced from arXiv: 2607.15849 by Dimitrios Karageorgiou, Efstratios Gavves, Ioannis Kompatsiaris, Symeon Papadopoulos.

Figure 1
Figure 1. Figure 1: While terminal points reside in the learned real manifold, no continuation from them exists inside it. Thus, they act as traps that force the generated trajectory to continue in unrealistic regions. By employing TANGO, we avoid terminal points at test time through noise guided optimization, thus keeping the generative process within the real manifold. The upper half of the figure visualizes the trajectory … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the test-time adaptation process for avoiding terminal points through noise guidance. The current output of the denoiser is criticized by moving a step forward and inspecting the properties of the predicted noise residual. Deviations from an isotropic Gaussian distribution are captured by a noise consistency objective. This objective is employed under a constrained optimization process, along w… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison against state-of-the-art approaches [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Failure cases. From left to right, frames indicative of four temporal video segments, of 15 s total duration, are presented. Two videos are considered, spanning the niche topics of exotic physics and nanobiology. While results are visually pleasant, technically, they are completely inaccurate. performance primarily by scaling up the model’s size. We report the respective results in Tab. 4. Overall, steerin… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

67 extracted references · 10 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2501.03575 (2025)

    Agarwal, N., Ali, A., Bala, M., Balaji, Y., Barker, E., Cai, T., Chattopadhyay, P., Chen, Y., Cui, Y., Ding, Y., et al.: Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575 (2025)

    Pith/arXiv arXiv 2025

  2. [2]

    arXiv preprint arXiv:2511.00062 (2025)

    Ali, A., Bai, J., Bala, M., Balaji, Y., Blakeman, A., Cai, T., Cao, J., Cao, T., Cha, E., Chao, Y.W., et al.: World simulation with video foundation models for physical ai. arXiv preprint arXiv:2511.00062 (2025)

    Pith/arXiv arXiv 2025

  3. [3]

    In: SIGGRAPH Asia 2024 Conference Papers

    Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Liu, G., Raj, A., et al.: Lumiere: A space-time diffusion model for video generation. In: SIGGRAPH Asia 2024 Conference Papers. pp. 1–11 (2024)

    2024

  4. [4]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 22563–22575 (2023)

    2023

  5. [5]

    Brooks, T., Peebles, B., Holmes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024),https://openai.com/index/video- generation-models-as-world-simulators/

    2024

  6. [6]

    arXiv preprint arXiv:2504.13074 (2025)

    Chen, G., Lin, D., Yang, J., Lin, C., Zhu, J., Fan, M., Zhang, H., Chen, S., Chen, Z., Ma, C., et al.: Skyreels-v2: Infinite-length film generative model. arXiv preprint arXiv:2504.13074 (2025)

    Pith/arXiv arXiv 2025

  7. [7]

    In: European Conference on Computer Vision

    Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. In: European Conference on Computer Vision. pp. 370–387. Springer (2024)

    2024

  8. [8]

    In: International Conference on Learning Representations (ICLR) (2023)

    Chung, H.W., Constant, N., Garcia, X., Roberts, A., Tay, Y., Narang, S., Firat, O.: Unimax: Fairer and more effective language sampling for large-scale multilingual pretraining. In: International Conference on Learning Representations (ICLR) (2023)

    2023

  9. [9]

    arXiv preprint arXiv:1907.06571 (2019)

    Clark, A., Donahue, J., Simonyan, K.: Adversarial video generation on complex datasets. arXiv preprint arXiv:1907.06571 (2019)

    Pith/arXiv arXiv 1907

  10. [10]

    IEEE transactions on pattern analysis and machine intelligence45(9), 10850–10869 (2023)

    Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. IEEE transactions on pattern analysis and machine intelligence45(9), 10850–10869 (2023)

    2023

  11. [11]

    arXiv preprint arXiv:2601.16914 (2026)

    Cui, J., Wu, J., Li, M., Yang, T., Li, X., Wang, R., Bai, A., Ban, Y., Hsieh, C.J.: Lol: Longer than longer, scaling video generation to hour. arXiv preprint arXiv:2601.16914 (2026)

    arXiv 2026

  12. [12]

    In: International Conference on Learning Representations (ICLR) (2024)

    Dao, T.: FlashAttention-2: Faster attention with better parallelism and work partitioning. In: International Conference on Learning Representations (ICLR) (2024)

    2024

  13. [13]

    In: International Conference on Learning Representations (ICLR) (2025)

    Deng, H., Pan, T., Diao, H., Luo, Z., Cui, Y., Lu, H., Shan, S., Qi, Y., Wang, X.: Autoregressive video generation without vector quantization. In: International Conference on Learning Representations (ICLR) (2025)

    2025

  14. [14]

    ACM Computing Surveys58(3), 1–38 (2025)

    Ding, J., Zhang, Y., Shang, Y., Zhang, Y., Zong, Z., Feng, J., Yuan, Y., Su, H., Li, N., Sukiennik, N., et al.: Understanding world or predicting future? a comprehensive survey of world models. ACM Computing Surveys58(3), 1–38 (2025)

    2025

  15. [15]

    Journal of the Optical Society of America A4(12), 2379–2394 (1987) Test-Time Noise Guided Adaptation for Realistic Video Generation 17

    Field, D.J.: Relations between the statistics of natural images and the response properties of cortical cells. Journal of the Optical Society of America A4(12), 2379–2394 (1987) Test-Time Noise Guided Adaptation for Realistic Video Generation 17

    1987

  16. [16]

    IEEE Transactions on Acoustics, Speech, and Signal Processing22(3), 207–217 (1974)

    Gray, A., Markel, J.: A spectral-flatness measure for studying the autocorrelation method of linear prediction of speech analysis. IEEE Transactions on Acoustics, Speech, and Signal Processing22(3), 207–217 (1974)

    1974

  17. [17]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Henschel, R., Khachatryan, L., Poghosyan, H., Hayrapetyan, D., Tadevosyan, V., Wang, Z., Navasardyan, S., Shi, H.: Streamingt2v: Consistent, dynamic, and extendable long video generation from text. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 2568–2577 (2025)

    2025

  18. [18]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  19. [19]

    Advances in neural information processing systems35, 8633–8646 (2022)

    Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. Advances in neural information processing systems35, 8633–8646 (2022)

    2022

  20. [20]

    International Conference on Learning Representations (ICLR) (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. International Conference on Learning Representations (ICLR) (2022)

    2022

  21. [21]

    IEEE transactions on pattern analysis and machine intelligence43(5), 1562–1577 (2019)

    Huang, L., Zhao, X., Huang, K.: Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE transactions on pattern analysis and machine intelligence43(5), 1562–1577 (2019)

    2019

  22. [22]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2025)

    Huang, X., Li, Z., He, G., Zhou, M., Shechtman, E.: Self forcing: Bridging the train-test gap in autoregressive video diffusion. In: Advances in Neural Information Processing Systems (NeurIPS) (2025)

    2025

  23. [23]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Huang, Z., He, Y., Yu, J., Zhang, F., Si, C., Jiang, Y., Zhang, Y., Wu, T., Jin, Q., Chanpaisit, N., et al.: Vbench: Comprehensive benchmark suite for video generative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21807–21818 (2024)

    2024

  24. [24]

    In: International Conference on Machine Learning (ICML) (2026)

    Jang, S., Ki, T., Jo, J., Xie, S., Yoon, J., Hwang, S.J.: Self-refining video sampling. In: International Conference on Machine Learning (ICML) (2026)

    2026

  25. [25]

    In: Uncertainty in Artificial Intelligence (UAI) (2025)

    Jazbec, M., Wong-Toi, E., Xia, G., Zhang, D., Nalisnick, E., Mandt, S.: Generative uncertainty in diffusion models. In: Uncertainty in Artificial Intelligence (UAI) (2025)

    2025

  26. [26]

    Advances in Neural Information Processing Systems37, 89834–89868 (2024)

    Kim, J., Kang, J., Choi, J., Han, B.: Fifo-diffusion: Generating infinite videos from text without training. Advances in Neural Information Processing Systems37, 89834–89868 (2024)

    2024

  27. [27]

    In: International Conference on Learning Representations (ICLR) (2025)

    Kim, S., Kim, M., Park, D.: Test-time alignment of diffusion models without reward over-optimization. In: International Conference on Learning Representations (ICLR) (2025)

    2025

  28. [28]

    Advances in neural information processing systems34, 21696–21707 (2021)

    Kingma, D., Salimans, T., Poole, B., Ho, J.: Variational diffusion models. Advances in neural information processing systems34, 21696–21707 (2021)

    2021

  29. [29]

    In: International Conference on Learning Representations (ICLR) (2014)

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: International Conference on Learning Representations (ICLR) (2014)

    2014

  30. [30]

    In: International Conference on Machine Learning (ICML) (2024)

    Kondratyuk, D., Yu, L., Gu, X., Lezama, J., Huang, J., Schindler, G., Hornung, R., Birodkar, V., Yan, J., Chiu, M.C., et al.: Videopoet: A large language model for zero-shot video generation. In: International Conference on Machine Learning (ICML) (2024)

    2024

  31. [31]

    arXiv preprint arXiv:2412.03603 (2024)

    Kong, W., Tian, Q., Zhang, Z., Min, R., Dai, Z., Zhou, J., Xiong, J., Li, X., Wu, B., Zhang, J., et al.: Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603 (2024)

    Pith/arXiv arXiv 2024

  32. [32]

    arXiv preprint arXiv:2510.09212 (2025) 18 Karageorgiou et al

    Li, W., Pan, W., Luan, P.C., Gao, Y., Alahi, A.: Stable video infinity: Infinite-length video generation with error recycling. arXiv preprint arXiv:2510.09212 (2025) 18 Karageorgiou et al

    arXiv 2025

  33. [33]

    In: International Conference on Learning Representations (ICLR) (2023)

    Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: International Conference on Learning Representations (ICLR) (2023)

    2023

  34. [34]

    In: International Conference on Learning Representations (ICLR) (2026)

    Liu, K., Hu, W., Xu, J., Shan, Y., Lu, S.: Rolling forcing: Autoregressive long video diffusion in real time. In: International Conference on Learning Representations (ICLR) (2026)

    2026

  35. [35]

    In: International Conference on Learning Representations (ICLR) (2019)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (ICLR) (2019)

    2019

  36. [36]

    Advances in Neural Information Processing Systems37, 131434–131455 (2024)

    Lu, Y., Liang, Y., Zhu, L., Yang, Y.: Freelong: Training-free long video generation with spectralblend temporal attention. Advances in Neural Information Processing Systems37, 131434–131455 (2024)

    2024

  37. [37]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Lu, Y., Zeng, Y., Li, H., Ouyang, H., Wang, Q., Cheng, K.L., Zhu, J., Cao, H., Zhang, Z., Zhu, X., et al.: Reward forcing: Efficient streaming video generation with rewarded distribution matching distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 34385–34397 (2026)

    2026

  38. [38]

    Electronics letters45(23), 1195–1196 (2009)

    Madhu, N.: Note on measures for spectral flatness. Electronics letters45(23), 1195–1196 (2009)

    2009

  39. [39]

    arXiv preprint arXiv:2501.08219 (2025)

    Maliakel, P.J., Ilager, S., Brandic, I.: Investigating energy efficiency and perfor- mance trade-offs in llm inference across tasks and dvfs settings. arXiv preprint arXiv:2501.08219 (2025)

    arXiv 2025

  40. [40]

    Advances in Neural Information Processing Systems37, 63858–63872 (2024)

    Miao, Y., Zhu, Y., Yu, L., Zhu, J., Gao, X.S., Dong, Y.: T2vsafetybench: Evaluating the safety of text-to-video generative models. Advances in Neural Information Processing Systems37, 63858–63872 (2024)

    2024

  41. [41]

    Advances in Neural Information Processing Systems36, 50358–50376 (2023)

    Muennighoff, N., Rush, A., Barak, B., Le Scao, T., Tazi, N., Piktus, A., Pyysalo, S., Wolf, T., Raffel, C.A.: Scaling data-constrained language models. Advances in Neural Information Processing Systems36, 50358–50376 (2023)

    2023

  42. [42]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4195–4205 (2023)

    2023

  43. [43]

    arXiv preprint arXiv:2310.15169 (2023)

    Qiu, H., Xia, M., Zhang, Y., He, Y., Wang, X., Shan, Y., Liu, Z.: Freenoise: Tuning- free longer video diffusion via noise rescheduling. arXiv preprint arXiv:2310.15169 (2023)

    Pith/arXiv arXiv 2023

  44. [44]

    arXiv preprint arXiv:2510.07940 (2025)

    Qu, L., Wang, Z., Zheng, N., Wang, W., Nie, L., Chua, T.S.: Ttom: Test-time optimization and memorization for compositional video generation. arXiv preprint arXiv:2510.07940 (2025)

    arXiv 2025

  45. [45]

    In: International Conference on Medical image computing and computer-assisted intervention

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234–241. Springer (2015)

    2015

  46. [46]

    In: International Conference on Learning Representations (ICLR) (2023)

    Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text- video data. In: International Conference on Learning Representations (ICLR) (2023)

    2023

  47. [47]

    In: International Conference on Learning Representations (ICLR) (2021)

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: International Conference on Learning Representations (ICLR) (2021)

    2021

  48. [48]

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- basedgenerativemodelingthroughstochasticdifferentialequations.In:International Conference on Learning Representations (ICLR) (2021)

    2021

  49. [49]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Sun, P., Cao, J., Jiang, Y., Yuan, Z., Bai, S., Kitani, K., Luo, P.: Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20993–21002 (2022) Test-Time Noise Guided Adaptation for Realistic Video Generation 19

    2022

  50. [50]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Tulyakov, S., Liu, M.Y., Yang, X., Kautz, J.: Mocogan: Decomposing motion and content for video generation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1526–1535 (2018)

    2018

  51. [51]

    In: Deep Generative Models for Highly Structured Data, ICLR 2019 Workshop (2019)

    Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., Gelly, S.: Fvd: A new metric for video generation. In: Deep Generative Models for Highly Structured Data, ICLR 2019 Workshop (2019)

    2019

  52. [52]

    In: International Conference on Learning Representations (ICLR) (2023)

    Villegas, R., Babaeizadeh, M., Kindermans, P.J., Moraldo, H., Zhang, H., Saffar, M.T., Castro, S., Kunze, J., Erhan, D.: Phenaki: Variable length video generation from open domain textual description. In: International Conference on Learning Representations (ICLR) (2023)

    2023

  53. [53]

    arXiv preprint arXiv:2503.20314 (2025)

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.W., Chen, D., Yu, F., Zhao, H., Yang, J., et al.: Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314 (2025)

    Pith/arXiv arXiv 2025

  54. [54]

    arXiv preprint arXiv:2305.18264 (2023)

    Wang, F.Y., Chen, W., Song, G., Ye, H.J., Liu, Y., Li, H.: Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264 (2023)

    Pith/arXiv arXiv 2023

  55. [55]

    Advances in neural information processing systems35, 24824–24837 (2022)

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems35, 24824–24837 (2022)

    2022

  56. [56]

    arXiv preprint arXiv:2602.05871 (2026)

    Xiang, X., Duan, Z., Zhang, G., Zhang, H., Gao, Z., Wu, J., Zhang, S., Wang, T., Fan, Q., Guo, C.: Pathwise test-time correction for autoregressive long video generation. arXiv preprint arXiv:2602.05871 (2026)

    arXiv 2026

  57. [57]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2025)

    Xie, D., Xu, Z., Hong, Y., Tan, H., Liu, D., Liu, F., Kaufman, A., Zhou, Y.: Progressive autoregressive video diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (2025)

    2025

  58. [58]

    ACM Comput

    Xing, Z., Feng, Q., Chen, H., Dai, Q., Hu, H., Xu, H., Wu, Z., Jiang, Y.G.: A survey on video diffusion models. ACM Comput. Surv.57(2) (Nov 2024).https: //doi.org/10.1145/3696415,https://doi.org/10.1145/3696415

    doi:10.1145/3696415 2024

  59. [59]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288–5296 (2016)

    2016

  60. [60]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Xue, H., Hang, T., Zeng, Y., Sun, Y., Liu, B., Yang, H., Fu, J., Guo, B.: Advancing high-resolution video-language representation with large-scale video transcriptions. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5036–5045 (2022)

    2022

  61. [61]

    arXiv preprint arXiv:2104.10157 (2021)

    Yan, W., Zhang, Y., Abbeel, P., Srinivas, A.: Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157 (2021)

    Pith/arXiv arXiv 2021

  62. [62]

    ACM computing surveys56(4), 1–39 (2023)

    Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Zhang, W., Cui, B., Yang, M.H.: Diffusion models: A comprehensive survey of methods and applications. ACM computing surveys56(4), 1–39 (2023)

    2023

  63. [63]

    In: International Conference on Learning Representations (ICLR) (2026)

    Yang, S., Huang, W., Chu, R., Xiao, Y., Zhao, Y., Wang, X., Li, M., Xie, E., Chen, Y., Lu, Y., et al.: Longlive: Real-time interactive long video generation. In: International Conference on Learning Representations (ICLR) (2026)

    2026

  64. [64]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yin, T., Zhang, Q., Zhang, R., Freeman, W.T., Durand, F., Shechtman, E., Huang, X.: From slow bidirectional to fast autoregressive video diffusion models. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22963–22974 (2025)

    2025

  65. [65]

    20 Karageorgiou et al

    Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A.G., Yang, M.H., Hao, Y., Essa, I., et al.: Magvit: Masked generative video transformer. 20 Karageorgiou et al. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10459–10469 (2023)

    2023

  66. [66]

    arXiv preprint arXiv:2512.11203 (2025)

    Yu, Z., Hayakawa, A., Ishii, M., Yu, Q., Shibuya, T., Zhang, J., Mitsufuji, Y.: Autorefiner:Improvingautoregressivevideodiffusionmodelsviareflectiverefinement over the stochastic sampling path. arXiv preprint arXiv:2512.11203 (2025)

    arXiv 2025

  67. [67]

    Zhang, Z., Chang, S., He, Y., Han, Y., Tang, J., Wang, F., Zhuang, B.: Blockvid: Block diffusion for high-quality and consistent minute-long video generation. arXiv preprint arXiv:2511.22973 (2025) Test-Time Noise Guided Adaptation for Realistic Autoregressive Video Generation Supplementary Material A1 Flow Matching Adaptation In the main paper we formula...

    Pith/arXiv arXiv 2025