Pith. sign in

REVIEW 2 major objections 5 minor 2 cited by

Scoring partially denoised latents during video diffusion enables efficient search that raises quality while cutting runtime by up to 79 percent.

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 · grok-4.5

2026-07-14 21:08 UTC pith:ZK3YJB4O

load-bearing objection Solid systems paper: intermediate latent rewards make stronger video-diffusion search practical, with clear quality/runtime gains and one acknowledged approximation. the 2 major comments →

arxiv 2603.14526 v2 pith:ZK3YJB4O submitted 2026-03-15 cs.CV

LatSearch: Latent Reward-Guided Search for Faster Inference-Time Scaling in Video Diffusion

Zengqun Zhao , Ziquan Liu , Yu Cao , Shaogang Gong , Zhensong Zhang , Jifei Song , Jiankang Deng , Ioannis Patras This is my paper
classification cs.CV
keywords video generationdiffusion modelslatent reward modelsinference-time scalingtext-to-videoreward-guided searchRGRP
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.

Prior inference-time scaling for video diffusion either tweaks only the starting noise or scores fully decoded videos, so feedback arrives late, errors accumulate, and strong search becomes too expensive. This paper establishes that a latent reward model can score noisy intermediate latents for visual quality, motion quality, and text alignment at any timestep, supplying early process-level signals. Building on those scores, LatSearch resamples candidate trajectories by reward-normalized probabilities and finally prunes to the single highest cumulative-reward path. On VBench-2.0 the method consistently improves creativity, commonsense, controllability, human fidelity, and physics over the Wan2.1 baseline, matching or beating earlier search methods while using far less compute. A sympathetic reader cares because intermediate latent feedback finally makes stronger, more sample-efficient search practical for video generation.

Core claim

A latent reward model that evaluates partially denoised video latents at arbitrary timesteps supplies intermediate, informative feedback that makes Reward-Guided Resampling and Pruning practical; the resulting LatSearch procedure improves generation quality across VBench-2.0 dimensions relative to Wan2.1 and achieves comparable or better quality than prior inference-time scaling methods while reducing runtime by up to 79 percent.

What carries the argument

LatSearch with Reward-Guided Resampling and Pruning (RGRP): a transformer reward model scores intermediate latents for visual quality, motion quality, and text alignment; candidates are then probabilistically resampled by those normalized scores (to avoid over-trusting an imperfect model) and finally pruned to the trajectory with highest cumulative reward.

Load-bearing premise

Cosine similarity between a partially denoised latent and the final clean latent is a faithful enough proxy for semantic contribution that simply scaling the finished video’s reward by that scalar yields usable training targets for intermediate scoring.

What would settle it

Replace cosine-similarity credit assignment with uniform or random weighting when building the latent-reward training set; if LatSearch then loses its VBench-2.0 gains while still using the same RGRP schedule, the similarity proxy is load-bearing and the central claim fails.

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

If this is right

  • Video quality rises across creativity, commonsense, controllability, human fidelity, and physics on VBench-2.0 relative to the Wan2.1 baseline.
  • Runtime can fall by up to 79 percent versus evolutionary search while matching or exceeding quality.
  • The same latent-space procedure extends to larger backbones without architecture changes.
  • Early pruning of low-reward trajectories reduces wasted denoising steps and limits error accumulation.
  • Once intermediate latent rewards replace full-video decoding, stronger search algorithms become computationally feasible.

Where Pith is reading between the lines

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

  • The same intermediate-scoring idea could transfer to image diffusion or audio-video generation by swapping only the text-alignment objective.
  • A learned temporal contribution estimator could replace cosine similarity and further raise latent-reward accuracy.
  • If latent rewards become reliable enough, continuous sequential-Monte-Carlo-style search could replace the fixed scoring schedule.
  • Controllability gains suggest latent guidance could support interactive editing or hard constraints without retraining the generator.

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

2 major / 5 minor

Summary. The paper introduces LatSearch, an inference-time scaling method for video diffusion models that uses a latent reward model (built on Qwen2-VL-3B with VQ/MQ/TA query tokens) to score partially denoised latents at arbitrary timesteps for visual quality, motion quality, and text alignment. Latent targets are obtained by scaling video-level rewards by cosine similarity between intermediate and final latents (Eqs. 7–8). Building on this, Reward-Guided Resampling and Pruning (RGRP) maintains multiple candidate trajectories, resamples them via reward-normalized multinomial probabilities with uniqueness pruning (Eqs. 16–17), and finally retains the highest cumulative-reward trajectory. Evaluated on VBench-2.0 with Wan2.1-1.3B (and 14B), LatSearch improves average scores over the baseline (up to +3.35) and matches or exceeds prior search methods (VideoReward, EvoSearch) while cutting runtime by up to 79%.

Significance. If the empirical gains hold, the work meaningfully advances inference-time scaling for video diffusion by making intermediate latent feedback practical, thereby enabling stronger search algorithms without the prohibitive cost of full decoding. The combination of a process-level reward model, probabilistic resampling that mitigates reward-model error, and extensive controlled ablations (search budget N, credit assignment, preference loss, temperature, scoring schedule), plus a human preference study (75% average preference) and runtime breakdowns, constitutes a solid, reproducible contribution. The method is model-agnostic and already demonstrated on a larger backbone, which strengthens its potential impact on controllable, efficient video generation.

major comments (2)
  1. Table 4 reports a baseline whose per-dimension scores (Controllability 21.79, Human Fidelity 89.74, Physics 39.96, Average 52.58) do not match the Wan2.1-1.3B baseline of Table 1 (21.99/82.11/45.98/51.90) nor the 14B baseline of Table 3; the deltas claimed for RGRP and preference loss are therefore computed against an inconsistent reference and cannot be trusted without correction.
  2. Section 3.2 (Eqs. 7–8) and Limitations §6 treat cosine similarity between zt and z0 as a scalar proxy for semantic contribution when constructing latent targets. While Table 5 and Appendix C.1 ablate alternatives and show cosine is best, the paper never measures how well st correlates with true quality contribution on held-out trajectories or under distribution shift; a quantitative correlation or oracle experiment would make the central credit-assignment assumption less of an untested approximation.
minor comments (5)
  1. Figure 1 caption and prompt text contain typos (“changfrom”, missing spaces); clean for camera-ready.
  2. In §4.1 the statement “for each noise prior, we generate 3,860 videos” is ambiguous; clarify whether this is the full VBench-2.0 prompt set or a subset.
  3. Notation for cumulative weight c_i^(t) (Eq. 18) is introduced without an explicit recurrence; a one-line definition would improve readability.
  4. Appendix Algorithm 1 uses r^d for both video-level and similarity-scaled targets; distinguish the two symbols to avoid confusion with the main-text ˜r_t.
  5. Table 6 reports slightly different absolute times from Table 1 for the same methods; reconcile the measurement protocol (hardware, batching, warm-up).

Circularity Check

1 steps flagged

No load-bearing circularity; cosine credit assignment is an explicit approximation for supervision, not a forced prediction of the reported VBench gains.

specific steps
  1. other [Section 3.2, Latent Reward Data Construction, Eqs. 7–8]
    "we ground video-level rewards to intermediate latents by measuring how much an intermediate latent zt “contributes” to the final clear latent z0 via its similarity to the clear latent. We define a cosine-based similarity, rescaled to [0,1]: st = 1/2 (1 + ⟨zt,z0⟩ / (∥zt∥2 ∥z0∥2)) ∈ [0,1]. ... Finally, we assign latent-level targets by crediting each dimension of the video-level reward proportionally to st: ˜rt = st · r"

    The regression targets for the latent reward model are defined by construction as the product of final video rewards and cosine similarity to the clean latent. This is a deliberate engineering choice for process supervision rather than a claimed first-principles derivation, and downstream claims are evaluated on independent VBench-2.0 metrics, so the circularity is only definitional of the training labels and does not force the reported generation improvements.

full rationale

The paper is an empirical methods contribution for inference-time search in video diffusion. Its central claims are measured on the independent VBench-2.0 suite (and a human preference study) using non-overlapping prompts for reward-model training. The only self-referential construction is the similarity-grounded target ˜rt = st · r (Eqs. 7–8), which defines the regression labels for the latent reward model; this is openly treated as an approximation (Limitations §6), ablated against Uniform/Exponential/L2 alternatives (Table 5, Appendix C.1), and further regularized by preference loss on the original video-level labels. RGRP (Eqs. 16–18) explicitly resamples probabilistically to reduce over-reliance on imperfect intermediate scores. No equation forces the reported quality or runtime gains by construction, no uniqueness theorem is imported from overlapping authors, and no fitted parameter is renamed as an out-of-sample prediction. The evaluation loop is therefore external and non-circular. Score 1 only for the minor definitional character of the credit-assignment labels themselves.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 2 invented entities

The paper is an empirical systems contribution. Its load-bearing free parameters are the usual search and training hyper-parameters chosen by ablation; its main ad-hoc axiom is the cosine-similarity credit assignment that turns final video rewards into intermediate targets; the invented entities are the latent reward architecture and the RGRP procedure themselves. No new physical or mathematical primitives are postulated.

free parameters (6)
  • search budget N
    Number of parallel latent trajectories; chosen by ablation (N=4/6/8) and fixed at 6 for main results; directly controls quality–cost trade-off.
  • scoring schedule S
    Timesteps at which the latent reward is queried (default {10,15,20}); selected by ablation over several mid-range schedules.
  • resampling temperature τ
    Softmax temperature converting rewards to multinomial probabilities; default 1.0, ablated in {0.5,1,2}.
  • noise mixing coefficient η
    Controls diversity of initial candidate set around base Gaussian noise; free design choice.
  • regression/preference loss weights λ
    Both set to 1.0; equal weighting is a free modeling decision.
  • credit-assignment function (cosine similarity)
    Choice among Uniform/Exponential/L2/Cosine; cosine selected after ablation and becomes the default target construction.
axioms (4)
  • ad hoc to paper Cosine similarity between intermediate latent zt and final clean latent z0 is a valid scalar proxy for the contribution of zt to final video quality, allowing linear scaling of video-level rewards.
    Introduced in §3.2 Eqs. 7–8; no independent theoretical justification beyond empirical ablation against three alternatives.
  • domain assumption External video-level verifiers (or human annotations) for VQ/MQ/TA supply ground-truth that can be transferred to latents via the similarity scalar.
    Standard in reward-model literature; invoked when constructing D_latent.
  • domain assumption Multinomial resampling with uniqueness followed by cumulative-weight pruning yields a useful importance-sampling approximation for latent trajectories.
    Inspired by SMC; used throughout §3.3 without convergence proof (acknowledged in Limitations).
  • domain assumption The UniPC sampler and classifier-free guidance with the stated schedules correctly implement the reverse diffusion process of Wan2.1.
    Standard diffusion machinery taken from prior literature and the Wan2.1 paper.
invented entities (2)
  • Latent reward model (Qwen2-VL-3B backbone with [VQ]/[MQ]/[TA] query tokens and 3-D patch encoder) no independent evidence
    purpose: Score partially denoised latents at arbitrary timesteps for visual quality, motion quality and text alignment without decoding.
    New architecture/training recipe for intermediate video latents; independent_evidence false because its utility is demonstrated only inside this paper’s search loop.
  • Reward-Guided Resampling and Pruning (RGRP) no independent evidence
    purpose: Convert latent reward scores into a practical multi-trajectory search that balances exploration (probabilistic resampling) and efficiency (uniqueness + final prune).
    Named procedure introduced in §3.3; no external validation outside the reported experiments.

pith-pipeline@v1.1.0-grok45 · 27456 in / 3157 out tokens · 47665 ms · 2026-07-14T21:08:30.433337+00:00 · methodology

0 comments
read the original abstract

The recent success of inference-time scaling in large language models has inspired similar explorations in video diffusion. In particular, motivated by the existence of "golden noise" that enhances video quality, prior work has attempted to improve inference by optimising or searching for better initial noise. However, these approaches have notable limitations: they either rely on priors imposed at the beginning of noise sampling or on rewards evaluated only on the denoised and decoded videos. This leads to error accumulation, delayed and sparse reward signals, and prohibitive computational cost, which prevents the use of stronger search algorithms. Crucially, stronger search algorithms are precisely what could unlock substantial gains in controllability, sample efficiency and generation quality for video diffusion, provided their computational cost can be reduced. To fill in this gap, we enable efficient inference-time scaling for video diffusion through latent reward guidance, which provides intermediate, informative and efficient feedback along the denoising trajectory. We introduce a latent reward model that scores partially denoised latents at arbitrary timesteps with respect to visual quality, motion quality, and text alignment. Building on this model, we propose LatSearch, a novel inference-time search mechanism that performs Reward-Guided Resampling and Pruning (RGRP). In the resampling stage, candidates are sampled according to reward-normalised probabilities to reduce over-reliance on the reward model. In the pruning stage, applied at the final scheduled step, only the candidate with the highest cumulative reward is retained, improving both quality and efficiency. We evaluate LatSearch on the VBench-2.0 benchmark and demonstrate that it consistently improves video generation across multiple evaluation dimensions compared to the baseline Wan2.1 model.

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. CachedSearch: Training-Free Cached Exploration for Test-Time Search in Video Diffusion

    cs.AI 2026-07 conditional novelty 6.0

    Using cached drafts to select the winner and regenerating only that winner captures 94.7% of best-of-8 search gain at 63% of the cost.

  2. Stream-T1: Test-Time Scaling for Streaming Video Generation

    cs.CV 2026-05 unverdicted novelty 6.0

    Stream-T1 is a test-time scaling framework for streaming video generation using scaled noise propagation from history, reward pruning across short and long windows, and feedback-guided memory sinking to improve tempor...

Reference graph

Works this paper leans on

64 extracted references · 26 linked inside Pith · cited by 2 Pith papers

  1. [1]

    arXiv preprint arXiv:2501.03575 (2025) 2

    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) 2

    Pith/arXiv arXiv 2025

  2. [2]

    In: ICLR (2025) 2, 4

    Ban, Y., Wang, R., Zhou, T., Gong, B., Hsieh, C.J., Cheng, M.: The crystal ball hypothesis in diffusion models: Anticipating object positions from initial noise. In: ICLR (2025) 2, 4

    2025

  3. [3]

    In: CVPR

    Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR. pp. 22669–22679 (2023) 4

    2023

  4. [4]

    In: ECCV

    Bhagat, S., Uppal, S., Yin, Z., Lim, N.: Disentangling multiple features in video sequences using gaussian processes in variational autoencoders. In: ECCV. pp. 102–117. Springer (2020) 4

    2020

  5. [5]

    arXiv preprint arXiv:2311.15127 (2023) 4

    Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) 4

    Pith/arXiv arXiv 2023

  6. [6]

    NeurIPS35, 31769–31781 (2022) 4

    Brooks, T., Hellsten, J., Aittala, M., Wang, T.C., Aila, T., Lehtinen, J., Liu, M.Y., Efros, A., Karras, T.: Generating long videos of dynamic scenes. NeurIPS35, 31769–31781 (2022) 4

    2022

  7. [7]

    In: CVPR

    Burgert, R., Xu, Y., Xian, W., Pilarski, O., Clausen, P., He, M., Ma, L., Deng, Y., Li, L., Mousavi, M., et al.: Go-with-the-flow: Motion-controllable video diffusion models using real-time warped noise. In: CVPR. pp. 13–23 (2025) 2, 5

    2025

  8. [8]

    In: ICCV

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV. pp. 9650– 9660 (2021) 5

    2021

  9. [9]

    In: ICLR (2024) 2, 5

    Chang, P., Tang, J., Gross, M., Azevedo, V.C.: How i warped your noise: a temporally-correlated noise prior for diffusion models. In: ICLR (2024) 2, 5

    2024

  10. [10]

    arXiv preprint arXiv:2310.19512 (2023) 4

    Chen, H., Xia, M., He, Y., Zhang, Y., Cun, X., Yang, S., Xing, J., Liu, Y., Chen, Q., Wang, X., et al.: Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512 (2023) 4

    Pith/arXiv arXiv 2023

  11. [11]

    In: CVPR

    Chen, S., Xu, M., Ren, J., Cong, Y., He, S., Xie, Y., Sinha, A., Luo, P., Xiang, T., Perez-Rua, J.M.: Gentron: Diffusion transformers for image and video generation. In: CVPR. pp. 6441–6451 (2024) 4

    2024

  12. [12]

    arXiv preprint arXiv:2305.13840 (2023) 4

    Chen, W., Ji, Y., Wu, J., Wu, H., Xie, P., Li, J., Xia, X., Xiao, X., Lin, L.: Control- a-video: Controllable text-to-video diffusion models with motion prior and reward feedback learning. arXiv preprint arXiv:2305.13840 (2023) 4

    Pith/arXiv arXiv 2023

  13. [13]

    In: CVPR

    Dalal, K., Koceja, D., Xu, J., Zhao, Y., Han, S., Cheung, K.C., Kautz, J., Choi, Y., Sun, Y., Wang, X.: One-minute video generation with test-time training. In: CVPR. pp. 17702–17711 (2025) 2, 4 16

    2025

  14. [14]

    arXiv preprint arXiv:2412.14169 (2024) 4

    Deng, H., Pan, T., Diao, H., Luo, Z., Cui, Y., Lu, H., Shan, S., Qi, Y., Wang, X.: Autoregressive video generation without vector quantization. arXiv preprint arXiv:2412.14169 (2024) 4

    Pith/arXiv arXiv 2024

  15. [15]

    NeurIPS 34, 8780–8794 (2021) 6

    Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS 34, 8780–8794 (2021) 6

    2021

  16. [16]

    arXiv preprint arXiv:2412.15689 (2024) 4

    Ding, Z., Jin, C., Liu, D., Zheng, H., Singh, K.K., Zhang, Q., Kang, Y., Lin, Z., Liu, Y.: Dollar: Few-step video generation via distillation and latent reward optimization. arXiv preprint arXiv:2412.15689 (2024) 4

    Pith/arXiv arXiv 2024

  17. [17]

    Doucet, A., De Freitas, N., Gordon, N.: An introduction to sequential monte carlo methods.In:SequentialMonteCarlomethodsinpractice,pp.3–14.Springer(2001) 9

    2001

  18. [18]

    arXiv preprint arXiv:2503.19325 (2025) 4

    Gu, Y., Mao, W., Shou, M.Z.: Long-context autoregressive video modeling with next-frame prediction. arXiv preprint arXiv:2503.19325 (2025) 4

    Pith/arXiv arXiv 2025

  19. [19]

    In: ICLR (2024) 4

    Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In: ICLR (2024) 4

    2024

  20. [20]

    arXiv preprint arXiv:2505.17618 (2025) 2, 3, 5, 10, 21

    He, H., Liang, J., Wang, X., Wan, P., Zhang, D., Gai, K., Pan, L.: Scaling image and video generation via test-time evolutionary search. arXiv preprint arXiv:2505.17618 (2025) 2, 3, 5, 10, 21

    Pith/arXiv arXiv 2025

  21. [21]

    NeurIPS30 (2017) 5

    Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS30 (2017) 5

    2017

  22. [22]

    In: ICLR (2024) 4

    Hiranaka, A., Chen, S.F., Lai, C.H., Kim, D., Murata, N., Shibuya, T., Liao, W.H., Sun, S.H., Mitsufuji, Y.: Hero: Human-feedback efficient reinforcement learning for online diffusion model finetuning. In: ICLR (2024) 4

    2024

  23. [23]

    arXiv preprint arXiv:2207.12598 (2022) 6

    Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) 6

    Pith/arXiv arXiv 2022

  24. [24]

    NeurIPS31(2018) 4

    Hsieh, J.T., Liu, B., Huang, D.A., Fei-Fei, L.F., Niebles, J.C.: Learning to de- compose and disentangle representations for video prediction. NeurIPS31(2018) 4

    2018

  25. [25]

    In: CVPR

    Kara, O., Kurtkaya, B., Yesiltepe, H., Rehg, J.M., Yanardag, P.: Rave: Random- ized noise shuffling for fast and consistent video editing with diffusion models. In: CVPR. pp. 6507–6516 (2024) 2

    2024

  26. [26]

    In: ICCV

    Karras, J., Holynski, A., Wang, T.C., Kemelmacher-Shlizerman, I.: Dreampose: Fashion image-to-video synthesis via stable diffusion. In: ICCV. pp. 22623–22633 (2023) 2

    2023

  27. [27]

    arXiv preprint arXiv:2505.17561 (2025) 2, 3, 4, 5

    Kim, K., Kim, S.: Model already knows the best noise: Bayesian active noise selec- tion via attention in video diffusion model. arXiv preprint arXiv:2505.17561 (2025) 2, 3, 4, 5

    arXiv 2025

  28. [28]

    arXiv preprint arXiv:2412.03603 (2024) 4

    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) 4

    Pith/arXiv arXiv 2024

  29. [29]

    arXiv preprint arXiv:2505.19196 (2025) 3

    Liao, X., Wei, W., Qu, X., Cheng, Y.: Step-level reward for free in rl-based t2i diffusion model fine-tuning. arXiv preprint arXiv:2505.19196 (2025) 3

    Pith/arXiv arXiv 2025

  30. [30]

    arXiv preprint arXiv:2503.18942 (2025) 2, 4

    Liu, F., Wang, H., Cai, Y., Zhang, K., Zhan, X., Duan, Y.: Video-t1: Test-time scaling for video generation. arXiv preprint arXiv:2503.18942 (2025) 2, 4

    Pith/arXiv arXiv 2025

  31. [31]

    In: NeurIPS (2025) 3, 4, 5, 7, 10, 21 17

    Liu, J., Liu, G., Liang, J., Yuan, Z., Liu, X., Zheng, M., Wu, X., Wang, Q., Qin, W., Xia, M., et al.: Improving video generation with human feedback. In: NeurIPS (2025) 3, 4, 5, 7, 10, 21 17

    2025

  32. [32]

    Machine Intelligence Research pp

    Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. Machine Intelligence Research pp. 1–22 (2025) 11

    2025

  33. [33]

    arXiv preprint arXiv:2502.07508 (2025) 4

    Luo, Y., Zhao, X., Chen, M., Zhang, K., Shao, W., Wang, K., Wang, Z., You, Y.: Enhance-a-video: Better generated video for free. arXiv preprint arXiv:2502.07508 (2025) 4

    Pith/arXiv arXiv 2025

  34. [34]

    In: CVPR

    Ma, N., Tong, S., Jia, H., Hu, H., Su, Y.C., Zhang, M., Yang, X., Li, Y., Jaakkola, T., Jia, X., et al.: Scaling inference time compute for diffusion models. In: CVPR. pp. 2523–2534 (2025) 2, 3, 4, 5

    2025

  35. [35]

    arXiv preprint arXiv:2501.05484 (2025) 4

    Ma, Y., Chen, J., Di, D., Xie, Q., Fan, L., Chen, W., Gou, X., Zhao, N., Yang, X.: Tuning-free long video generation via global-local collaborative diffusion. arXiv preprint arXiv:2501.05484 (2025) 4

    Pith/arXiv arXiv 2025

  36. [36]

    In: CVPR

    Nam, H., Kim, J., Lee, D., Ye, J.C.: Optical-flow guided prompt optimization for coherent video generation. In: CVPR. pp. 7837–7846 (2025) 4

    2025

  37. [37]

    In: NeurIPS (2025) 2, 3, 5

    Oshima, Y., Suzuki, M., Matsuo, Y., Furuta, H.: Inference-time text-to-video align- ment with diffusion latent beam search. In: NeurIPS (2025) 2, 3, 5

    2025

  38. [38]

    NeurIPS35, 27730–27744 (2022) 8

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. NeurIPS35, 27730–27744 (2022) 8

    2022

  39. [39]

    arXiv preprint arXiv:2507.15728 (2025) 4

    Ouyang, W., Xiao, Z., Yang, D., Zhou, Y., Yang, S., Yang, L., Si, J., Pan, X.: Tokensgen: Harnessing condensed tokens for long video generation. arXiv preprint arXiv:2507.15728 (2025) 4

    Pith/arXiv arXiv 2025

  40. [40]

    In: ICCV

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV. pp. 4195–4205 (2023) 4

    2023

  41. [41]

    In: ICMR

    Peruzzo, E., Xu, D., Xu, X., Shi, H., Sebe, N.: Ragme: Retrieval augmented video generation for enhanced motion realism. In: ICMR. pp. 1081–1090 (2025) 4

    2025

  42. [42]

    arXiv preprint arXiv:2407.14041 (2024) 2, 4

    Qi, Z., Bai, L., Xiong, H., Xie, Z.: Not all noises are created equally: Diffusion noise selection and optimization. arXiv preprint arXiv:2407.14041 (2024) 2, 4

    Pith/arXiv arXiv 2024

  43. [43]

    In: ICLR (2023) 4

    Qiu, H., Xia, M., Zhang, Y., He, Y., Wang, X., Shan, Y., Liu, Z.: Freenoise: Tuning- free longer video diffusion via noise rescheduling. In: ICLR (2023) 4

    2023

  44. [44]

    In: ICML

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748–8763 (2021) 5

    2021

  45. [45]

    NeurIPS29(2016) 5

    Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Improved techniques for training gans. NeurIPS29(2016) 5

    2016

  46. [46]

    arXiv preprint arXiv:2501.06848 (2025) 3, 5

    Singhal, R., Horvitz, Z., Teehan, R., Ren, M., Yu, Z., McKeown, K., Ranganath, R.: A general framework for inference-time scaling and steering of diffusion models. arXiv preprint arXiv:2501.06848 (2025) 3, 5

    Pith/arXiv arXiv 2025

  47. [47]

    In: CVPR

    Skorokhodov, I., Tulyakov, S., Elhoseiny, M.: Stylegan-v: A continuous video gener- ator with the price, image quality and perks of stylegan2. In: CVPR. pp. 3626–3636 (2022) 4

    2022

  48. [48]

    arXiv preprint arXiv:2010.02502 (2020) 8

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) 8

    Pith/arXiv arXiv 2010

  49. [49]

    In: ICLR (2020) 6

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. In: ICLR (2020) 6

    2020

  50. [50]

    arXiv preprint arXiv:2508.06082 (2025) 4 18

    Sun, Y., Wu, J., Cao, Y., Xu, C., Wang, Y., Cao, W., Luo, D., Wang, C., Fu, Y.: Swiftvideo: A unified framework for few-step video generation through trajectory- distribution alignment. arXiv preprint arXiv:2508.06082 (2025) 4 18

    arXiv 2025

  51. [51]

    arXiv preprint arXiv:2503.20314 (2025) 4, 10

    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) 4, 10

    Pith/arXiv arXiv 2025

  52. [52]

    arXiv preprint arXiv:2308.06571 (2023) 4

    Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text- to-video technical report. arXiv preprint arXiv:2308.06571 (2023) 4

    Pith/arXiv arXiv 2023

  53. [53]

    arXiv preprint arXiv:2409.12191 (2024) 10

    Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., et al.: Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191 (2024) 10

    Pith/arXiv arXiv 2024

  54. [54]

    In: ECCV

    Wu, T., Si, C., Jiang, Y., Huang, Z., Liu, Z.: Freeinit: Bridging initialization gap in video diffusion models. In: ECCV. pp. 378–394 (2024) 2, 5, 10, 21

    2024

  55. [55]

    NeurIPS37, 76115–76138 (2024) 4

    Xiao, Z., Zhou, Y., Yang, S., Pan, X.: Video diffusion models are training-free motion interpreter and controller. NeurIPS37, 76115–76138 (2024) 4

    2024

  56. [56]

    arXiv preprint arXiv:2503.16400 (2025) 2, 3, 5

    Yang, H., Tang, F., Hu, M., Yin, Q., Li, Y., Liu, Y., Peng, Z., Gao, P., He, J., Ge, Z., et al.: Scalingnoise: Scaling inference-time search for generating infinite videos. arXiv preprint arXiv:2503.16400 (2025) 2, 3, 5

    Pith/arXiv arXiv 2025

  57. [57]

    In: ICLR (2024) 4

    Yang, Z., Teng, J., Zheng, W., Ding, M., Huang, S., Xu, J., Yang, Y., Hong, W., Zhang, X., Feng, G., et al.: Cogvideox: Text-to-video diffusion models with an expert transformer. In: ICLR (2024) 4

    2024

  58. [58]

    In: CVPR

    Yuan, H., Zhang, S., Wang, X., Wei, Y., Feng, T., Pan, Y., Zhang, Y., Liu, Z., Albanie, S., Ni, D.: Instructvideo: Instructing video diffusion models with human feedback. In: CVPR. pp. 6463–6474 (2024) 4

    2024

  59. [59]

    In: ICLR (2025) 2, 5, 10, 21

    Yuan, Y., Guo, Y., Wang, C., Zhang, W., Xu, H., Zhang, L.: Freqprior: Improving video diffusion models with frequency filtering gaussian noise. In: ICLR (2025) 2, 5, 10, 21

    2025

  60. [60]

    arXiv preprint arXiv:2501.07563 (2025) 2, 5

    Zhang, X., Duan, Z., Gong, D., Liu, L.: Training-free motion-guided video gen- eration with enhanced temporal consistency using motion consistency loss. arXiv preprint arXiv:2501.07563 (2025) 2, 5

    Pith/arXiv arXiv 2025

  61. [61]

    NeurIPS36, 49842–49869 (2023) 6, 8

    Zhao, W., Bai, L., Rao, Y., Zhou, J., Lu, J.: Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. NeurIPS36, 49842–49869 (2023) 6, 8

    2023

  62. [62]

    arXiv preprint arXiv:2503.21755 (2025) 10

    Zheng, D., Huang, Z., Liu, H., Zou, K., He, Y., Zhang, F., Zhang, Y., He, J., Zheng, W.S., Qiao, Y., et al.: Vbench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness. arXiv preprint arXiv:2503.21755 (2025) 10

    Pith/arXiv arXiv 2025

  63. [63]

    arXiv preprint arXiv:2411.09502 (2024) 2, 4

    Zhou, Z., Shao, S., Bai, L., Zhang, S., Xu, Z., Han, B., Xie, Z.: Golden noise for diffusion models: A learning framework. arXiv preprint arXiv:2411.09502 (2024) 2, 4

    Pith/arXiv arXiv 2024

  64. [64]

    arXiv preprint arXiv:2504.10044 (2025) 4 19 Appendix This appendix provides supplementary materials that support the main paper

    Zhu, B., Jiang, Y., Xu, B., Yang, S., Yin, M., Wu, Y., Sun, H., Wu, Z.: Aligning anime video generation with human feedback. arXiv preprint arXiv:2504.10044 (2025) 4 19 Appendix This appendix provides supplementary materials that support the main paper. We first present the detailed algorithms of the proposed latent reward model training and the latent re...

    Pith/arXiv arXiv 2025