Consistency-Preserving Diverse Video Generation

Runfa Blark Li; Truong Nguyen; Xinshuang Liu

REVIEW 4 major objections 3 minor 23 references

A consistency-filtered diversity gradient lets flow-matching video samplers increase batch diversity while keeping each video temporally coherent.

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

2026-08-02 22:54 UTC pith:SATOL7MI

load-bearing objection Plausible latent-space gradient regulation for diversifying video flow-matching batches, but the consistency proxy is never validated against the metric it is supposed to preserve, and the diversity claims are overstated. the 4 major comments →

arxiv 2602.15287 v2 pith:SATOL7MI submitted 2026-02-17 cs.CV

Consistency-Preserving Diverse Video Generation

Xinshuang Liu , Runfa Blark Li , Truong Nguyen This is my paper

classification cs.CV

keywords text-to-video generationflow matchingdiversitytemporal consistencyjoint samplinglatent-space guidancegradient regulationdeterminantal point processes

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.

The paper asks how to get more out of a fixed budget of text-to-video samples: when only a handful of videos can be generated per prompt, the batch should be as diverse as possible. Prior joint-sampling tricks that push samples apart in image space are expensive for video and tend to make the frames within a video drift apart. The paper's proposal is to keep the diversity push, but first split it into a component that would hurt a temporal-consistency objective and a component that would not, then drop only the harmful part. Both the diversity and consistency objectives are evaluated by small latent-space models, so the whole loop runs without decoding frames or backpropagating through the video decoder. On a state-of-the-art flow-matching text-to-video model, the result is diversity close to strong joint-sampling baselines with better temporal consistency and color naturalness.

Core claim

The paper claims that a diversity gradient for a batch of flow-matching video latents can be decomposed relative to a consistency gradient. Writing the diversity gradient as a component parallel to the consistency gradient plus a perpendicular component, the regulated update keeps the perpendicular part and keeps the parallel part only when it points in the same direction as the consistency objective; when it points opposite, it is discarded. Since the consistency gradient is the gradient of a latent frame-interpolation loss, this guarantees the diversity update never decreases that consistency objective to first order. The authors further show that both the diversity embedding and the consi

What carries the argument

The load-bearing object is the regulated diversity gradient: g_reg = g_perp + max(alpha, 0) g_c. Here g_c is the gradient of a latent temporal-consistency objective O_c, alpha is the projection of the diversity gradient g_d onto g_c, and g_perp is the leftover component. This one formula turns any unconstrained diversity push into one that cannot decrease the consistency objective to first order, and it does so with purely latent-space computations. The supporting machinery is a set of lightweight latent models—a video embedding model, a frame embedding model, and a frame-interpolation model—trained to mimic frozen video-space encoders and an interpolation-based consistency metric, so no dec

Load-bearing premise

The whole method rests on the latent frame-interpolation loss O_c being a faithful first-order proxy for the temporal consistency of the decoded video; if that proxy is misaligned with real frame coherence, the filter can throw away harmless diversity or keep updates that still hurt consistency.

What would settle it

Compute the rank correlation between O_c and the decoded frame-interpolation MSE used for evaluation across a set of flow-matching latents sampled at different ODE steps. If O_c does not rank videos the same way as the decoded metric, the consistency filter is not preserving what it claims to preserve.

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

If this is right

Joint sampling can be run on text-to-video models without decoder forward/backward passes, bringing the compute cost down to the level of the base sampler plus small latent models.
The method achieves Vendi diversity comparable to unregulated joint-sampling baselines while cutting temporal-consistency MSE roughly by a third in the reported setup.
The consistency filter preserves diversity because it discards only the anti-aligned component of the diversity gradient, not the whole gradient.
The approach transfers to new prompts without retraining the base generator; only the small latent models need training.

Where Pith is reading between the lines

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

A direct testable extension is to check whether the latent interpolation loss O_c correlates with decoded temporal-consistency scores on a per-sample basis; if it does, the method should transfer to other flow-matching video backbones with only the lightweight surrogate models retrained.
The projection rule suggests a general recipe: for any batch-level diversity objective, split the gradient against a quality gradient and keep only the non-harmful component. This could apply to image batches, multi-view generation, or any expensive-decoder setting.
The max(alpha, 0) rule is scale-sensitive: it keeps the entire parallel component when aligned, which may over-correct in early sampling steps. A softened or normalized variant would be a natural next step, though the paper does not explore it.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Plausible latent-space gradient regulation for diversifying video flow-matching batches, but the consistency proxy is never validated against the metric it is supposed to preserve, and the diversity claims are overstated.

read the letter

The core idea is simple and mostly sound: after adding a diversity gradient to a flow-matching ODE, project out the component that a latent interpolation model says would hurt temporal consistency. The first-order projection in Eqs. 15–16 is correct, and the paper correctly frames it as an extension of the authors' image-quality regulation. The latent surrogate models M_v/M_f/M_c are a legit engineering contribution — computing both objectives without decoder backprop is genuinely useful for video, where parallel decoding is expensive.

What the paper does well: the experiments include confidence intervals, ablations, and a sensible base model (Wan 2.1). The consistency regulation clearly helps: MSE drops from ~0.0028 to ~0.0019 relative to the diversity-only baselines, and CNI rises. The ablation in Table II supports the claim that the regulation is doing the work. The training of the surrogate models is also tested against simple baselines (latent-mean embeddings, neighbor-based interpolation), which is more than many papers do.

The soft spots are real and load-bearing. The stress-test note is right: Eq. 13 defines the consistency objective as the MSE of a latent interpolation model, but the paper never checks whether lowering that loss actually lowers the EDEN MSE reported in Table I. Figure 3(f) only shows that M_c beats simple interpolation in latent space; it says nothing about transfer to decoded video. Without that correlation analysis, the phrase "consistency-preserving" describes a property of a latent surrogate, not of the generated videos. At finite step sizes, even a direction orthogonal to g_c can change O_c through curvature, so the first-order argument is not a substitute for empirical validation.

Two further issues. First, the diversity claim is overstated: Vendi-f for the full method is 0.197(3) versus 0.207–0.208(3) for all three baselines — a clear drop, not "comparable." Vendi-v is comparable, but Vendi-f is not. Second, the baselines use spatial-mean latent embeddings, while the proposed method uses trained surrogate models. The proper controlled experiment is the same learned embeddings with and without consistency regulation, on equal footing — the ablation does this internally, but the comparison to DPP/Particle Guidance/DiverseFlow is not apples-to-apples. Also, the abstract says code is available at a URL; the full text says "Code will be released." That inconsistency needs fixing.

The circularity concern (surrogates trained to mimic the same encoders used for evaluation) is real but secondary, since the metrics are computed on real decoded video with the true encoders; the surrogates are only used for guidance.

Who is this for: anyone working on joint sampling for flow-matching video generators, or on gradient guidance with latent-space proxies. It deserves a serious referee, but the referee should demand the O_c-to-EDEN correlation analysis, an equal-footing baseline, and toned-down claims. If the proxy transfer can be established, this becomes a useful chapter in the diversity-consistency story. Right now it is a solid but incomplete manuscript.

Referee Report

4 major / 3 minor

Summary. The paper proposes a joint-sampling framework for flow-matching text-to-video models that aims to increase cross-video diversity while preserving within-video temporal consistency. Diversity is encouraged through a DPP objective computed on latent-space embeddings from learned models M_v and M_f; a consistency objective O_c based on a latent frame-interpolation model M_c is used to remove from the diversity gradient the components that would decrease O_c to first order (Eqs. 13–16). All guidance computations are performed in latent space, avoiding video decoding and decoder backpropagation during sampling. Experiments on Wan 2.1 t2v-1.3B report diversity close to strong joint-sampling baselines while improving temporal consistency (lower EDEN MSE) and color naturalness (higher CNI).

Significance. The core idea — regulate a diversity gradient with a latent-space consistency proxy so that diversity updates do not harm temporal coherence — is novel and practically motivated. The first-order derivation in Eqs. (15)–(16) is correct, and the paper includes confidence intervals, ablations, and code release, which are strengths. However, the central claim is only as strong as the surrogate O_c: the paper never shows that O_c tracks the reported EDEN MSE. In addition, the diversity claim is partly contradicted by the Vendi-f results, and the diversity objective is computed with embeddings trained to mimic the very encoders used for evaluation. These gaps make the current evidence insufficient to support the paper's headline conclusions, though they are addressable with additional experiments.

major comments (4)

[§III-A, Eqs. (13)–(16); §IV-A and Table I] The consistency-preservation guarantee is only for the latent surrogate O_c = -||x̂1 - M_c(x̂1)||², not for the EDEN MSE used to evaluate temporal consistency in Table I. The paper never measures the correlation between O_c and EDEN, nor shows that a first-order preservation of O_c transfers through the decoder to the actual video frames. Figure 3(f) demonstrates that M_c has lower interpolation loss than simple baselines, but it does not establish that optimizing O_c improves EDEN MSE. This is load-bearing because the claim 'our method substantially improves temporal consistency' is about decoded videos, while the mathematical argument in Eq. (16) concerns only the latent proxy. The first-order argument is also local; at finite step sizes, curvature can cause even a direction orthogonal to g_c to change O_c. Please add a correlation analysis between O_c and EDEN (e.g., scatter over samp
[§IV-B, Table I] The statement 'Our method attains comparable diversity to these baselines' is not supported for Vendi-f. The reported 95% confidence intervals are non-overlapping: Ours Vendi-f = 0.197(3) versus DPP/Particle Guidance/DiverseFlow = 0.207(3)–0.208(3). This is a statistically significant shortfall. The claim is only defensible for Vendi-v, where intervals overlap. Please either temper the claim to acknowledge the Vendi-f deficit, or provide evidence (e.g., human evaluation or an additional diversity metric) that the Vendi-f gap does not reflect a practically meaningful loss of diversity.
[§III-B, Eqs. (17)–(21); §IV-A evaluation metrics] The latent embedding models M_v and M_f are trained to reproduce the pairwise dot products of VideoPrism-B and CLIP embeddings (Eq. 17), which are exactly the encoders used to compute Vendi-v and Vendi-f. Consequently, the diversity objective is aligned with the evaluation metric by construction, while the baselines use unaligned latent mean features. This gives the proposed method an advantage in the diversity comparison that may not reflect gains in human-perceived diversity. The fairness of the comparison would be improved by reporting diversity on a metric not used in the objective, such as LPIPS over decoded frames, or by training the latent models on data from held-out prompts to test generalization.
[§IV-C, Table II] The ablation toggles consistency regulation as a binary component but does not include a strength-matched control. The improvement in MSE when ConsisReg is enabled could partly come from a simple reduction in the effective diversity-gradient magnitude rather than from the specific projection onto the consistency subspace. A comparison against a scaled-down diversity velocity (matched to give the same Vendi-v as Ours) would isolate the benefit of the consistency-aware projection and support the claim that the regulation preserves diversity while protecting consistency.

minor comments (3)

[§II-B, Eq. (2) vs. §III-A, Eq. (10)] The notation D(i,i') is used first for a generic pairwise difference in Eq. (2) and then redefined as the average of video and frame differences in Eq. (10). Please use a different symbol for one of them to avoid confusion.
[Abstract] The abstract says 'Code will be released' while the submission header includes a GitHub URL; please make these consistent.
[Figure 3] The panel captions (b)–(e) say 'Loss of Random Embedding: 0.69' etc., but the meaning of this baseline value and how it is scaled are not explained in the text. Please clarify.

Circularity Check

0 steps flagged

No significant circularity: the core derivation preserves a defined latent objective by construction and the reported gains are independent empirical measurements.

full rationale

The paper's main construction is a gradient regulation that removes the component of the diversity gradient anti-aligned with ∇O_c. By Eq. 15–16 this guarantees first-order non-decrease of O_c, but O_c is explicitly defined as a latent interpolation objective (Eq. 13), not as the EDEN MSE used in Table I. The claim that the method 'significantly improves temporal consistency' is therefore an empirical claim about a separate metric, not a derivation forced by the definition of O_c. Similarly, the latent embedding models M_v/M_f are trained to mimic VideoPrism-B and CLIP, which are also used for the Vendi evaluation; this is an intentional surrogate alignment, not a fitted input renamed as a prediction—the reported Vendi scores are obtained from the actual frozen encoders on decoded videos. The one self-citation ([16]) is used only for a normalization choice and for noting that the gradient-regulation idea is related to prior work; the regulation is fully specified in the current paper and does not rely on [16] for its correctness. No equation reduces to an input by construction, and no load-bearing claim is forced by a self-citation chain. The absence of a correlation study between O_c and EDEN MSE is a legitimate external-validity concern, but it is a correctness risk, not a circularity per the stated criteria.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 4 invented entities

The central claim rests on several trained latent-space components, hand-chosen loss weights, and domain assumptions about proxy objectives. The latent models are empirically validated within the paper but not released, and the key proxy-to-metric alignment (O_c to EDEN) is unmeasured.

free parameters (5)

lambda_s (embedding loss weight) = 10.0
Set by hand in Eq 21; no sensitivity analysis is reported.
lambda_t (diversity velocity strength) = not reported
Defined in Eq 3; the text says the normalized velocity u is scaled to match the norm of v_theta, but the scalar schedule is not given.
Implicit weights on L_p, L_reg,m, L_reg,p = 1.0 (implicit)
Eq 21 sets all non-lambda_s terms to weight 1 without ablation or justification.
Sparsification schedule for A_v/A_f = not specified
The text says the projection matrices are gradually sparsified during the middle 4,000 training steps, but the threshold and schedule are not stated.
Latent model architectures and training hyperparameters = not reported
M_v, M_f, and M_c are described as small convolutional networks, but architecture, optimizer, learning rate, batch size, and training details are omitted.

axioms (5)

domain assumption Adding a diversity velocity u to the flow-matching ODE (Eq 4) yields valid samples from a modified generative process.
The coupled-ODE sampler from [5,6] is adopted without proving that perturbing v_theta preserves sample quality; only metrics are checked.
domain assumption The extrapolated terminal latent hat_x1 = x_t + (1-t) v_theta(x_t,t) is a good predictor of the final latent at every intermediate step.
Used in Eqs 5-6 and 13; validated indirectly in Fig 3 but assumed across all sampling steps.
domain assumption Trained latent models M_v/M_f/M_c generalize from training videos to test prompts and to the full sampling trajectory.
Training uses 100 videos per prompt; no cross-prompt or cross-model generalization experiment is reported.
standard math First-order Taylor expansion is sufficient for the gradient regulation in Eqs 15-16.
The argument that removing anti-aligned components preserves consistency is first-order; ODE step sizes and curvature are not analyzed.
domain assumption Vendi score on VideoPrism-B/CLIP, EDEN interpolation MSE, and CNI are accepted ground-truth measures of diversity, temporal consistency, and color naturalness.
These metrics are used without human evaluation or additional corroborating metrics.

invented entities (4)

M_v (latent video embedding model) no independent evidence
purpose: Computes video-level diversity embeddings in latent space without decoding (Eq 5).
Validated on held-out test videos inside the paper (Fig 3d-e), but no released weights or code permit external replication or falsification.
M_f (latent frame embedding model) no independent evidence
purpose: Computes frame-level diversity embeddings in latent space (Eq 6).
Validated against CLIP frame embeddings inside the paper (Fig 3b-c), but no released weights or code.
M_c (latent frame interpolation model) no independent evidence
purpose: Computes the temporal-consistency objective O_c in latent space (Eq 13).
Compared against simple interpolation baselines inside the paper (Fig 3f), but its alignment with EDEN consistency is not established and no code is released.
Learned alignment matrices A_v/A_f no independent evidence
purpose: Maps latent and video embeddings into a shared projected space for the projection operation (Eqs 5-6).
Trained with sparsity regularization; no external handle exists outside the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 6553 in / 19191 out tokens · 183769 ms · 2026-08-02T22:54:27.513750+00:00 · methodology

0 comments

read the original abstract

Text-to-video generation is expensive, so only a few samples are typically produced per prompt. In this low-sample regime, maximizing the value of each batch requires high cross-video diversity. Recent methods improve diversity for image generation, but for videos they often degrade within-video temporal consistency and require costly backpropagation through a video decoder. We propose a joint-sampling framework for flow-matching video generators that improves batch diversity while preserving temporal consistency. Our approach applies diversity-driven updates and then removes only the components that would decrease a temporal-consistency objective. To avoid image-space gradients, we compute both objectives with lightweight latent-space models, avoiding video decoding and decoder backpropagation. Experiments on a state-of-the-art text-to-video flow-matching model show diversity close to strong joint-sampling baselines while substantially improving temporal consistency and color naturalness. Our code is available at https://github.com/XinshuangL/Diverse-Video.

Figures

Figures reproduced from arXiv: 2602.15287 by Runfa Blark Li, Truong Nguyen, Xinshuang Liu.

**Figure 1.** Figure 1: Joint video generation with enhanced cross-video diversity and preserved intra-video temporal consistency based on [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Figure 2: Illustration of training procedure for latent-space em [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Model behavior across flow-matching steps: (a) Extrapolation of [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 2 linked inside Pith

[1]

Videomage: Multi-subject and motion customization of text-to-video diffusion models,

C. Huang, Y . Wu, H. Chung, K. Chang, F. Yang, and Y . F. Wang, “Videomage: Multi-subject and motion customization of text-to-video diffusion models,” in CVPR, 2025

2025
[2]

Cogvideox: Text-to-video diffusion models with an expert transformer,

Z. Yang et al., “Cogvideox: Text-to-video diffusion models with an expert transformer,” inICLR, 2025

2025
[3]

Wonderworld: Interactive 3d scene generation from a single image,

H. Yu, H. Duan, C. Herrmann, W. T. Freeman, and J. Wu, “Wonderworld: Interactive 3d scene generation from a single image,” inCVPR, 2025

2025
[4]

SVG: 3d stereoscopic video generation via denoising frame matrix,

P. Dai et al., “SVG: 3d stereoscopic video generation via denoising frame matrix,” inICLR, 2025

2025
[5]

Particle guidance: Non-i.i.d. diverse sampling with diffusion models,

G. Corso, Y . Xu, V . D. Bortoli, R. Barzilay, and T. S. Jaakkola, “Particle guidance: Non-i.i.d. diverse sampling with diffusion models,” inICLR, 2024

2024
[6]

Diverseflow: Sample- efficient diverse mode coverage in flows,

M. M. Morshed and V . Boddeti, “Diverseflow: Sample- efficient diverse mode coverage in flows,” inCVPR, 2025

2025
[7]

Flow matching for generative modeling,

Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inICLR, 2023

2023
[8]

Generative adversarial nets,

I. J. Goodfellow et al., “Generative adversarial nets,” in NIPS, 2014

2014
[9]

Generat- ing videos with scene dynamics,

C. V ondrick, H. Pirsiavash, and A. Torralba, “Generat- ing videos with scene dynamics,” inNIPS, 2016

2016
[10]

Density estimation using real NVP,

L. Dinh, J. Sohl-Dickstein, and S. Bengio, “Density estimation using real NVP,” inICLR, 2017

2017
[11]

Videoflow: A conditional flow-based model for stochastic video generation,

M. Kumar et al., “Videoflow: A conditional flow-based model for stochastic video generation,” inICLR, 2020

2020
[12]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inNeurIPS, 2020

2020
[13]

Video diffusion models,

J. Ho, T. Salimans, A. A. Gritsenko, W. Chan, M. Norouzi, and D. J. Fleet, “Video diffusion models,” in NeurIPS, 2022. Data distribution 10 20 30 40 Noise distribution Steps Final latents Prediction & Decoding … … … Decoded videos Video latents … 50 (a) Extrapolation ofˆx1 =x t + (1−t)v θ(xt, t)from intermediate latent states and decoding them to video fr...

2022
[14]

Flow straight and fast: Learning to generate and transfer data with rectified flow,

X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” inICLR, 2023

2023
[15]

Wan: Open and advanced large-scale video generative models,

A. Wang et al., “Wan: Open and advanced large-scale video generative models,”CoRR, vol. abs/2503.20314, 2025

Pith/arXiv arXiv 2025
[16]

Importance- weighted non-iid sampling for flow matching models,

X. Liu, R. B. Li, S. Wei, and T. Nguyen, “Importance- weighted non-iid sampling for flow matching models,” arXiv preprint arXiv:2511.17812, 2025

Pith/arXiv arXiv 2025
[17]

Determinantal point pro- cesses for machine learning,

A. Kulesza, B. Taskar, et al., “Determinantal point pro- cesses for machine learning,”F oundations and Trends® in Machine Learning, vol. 5, no. 2–3, pp. 123–286, 2012

2012
[18]

VideoPrism: A foundational visual encoder for video understanding,

L. Zhao et al., “VideoPrism: A foundational visual encoder for video understanding,” inICML, 2024

2024
[19]

Learning transferable visual models from natural language supervision,

A. Radford et al., “Learning transferable visual models from natural language supervision,” inICML, 2021

2021
[20]

The vendi score: A diversity evaluation metric for machine learning,

D. Friedman and A. B. Dieng, “The vendi score: A diversity evaluation metric for machine learning,”Trans. Mach. Learn. Res., 2023

2023
[21]

EDEN: enhanced diffusion for high- quality large-motion video frame interpolation,

Z. Zhang et al., “EDEN: enhanced diffusion for high- quality large-motion video frame interpolation,” in CVPR, 2025

2025
[22]

Natural color image enhancement and evaluation algorithm based on human visual system,

K. Huang, Q. Wang, and Z. Wu, “Natural color image enhancement and evaluation algorithm based on human visual system,”Comput. Vis. Image Underst., vol. 103, no. 1, pp. 52–63, 2006

2006
[23]

Robust multi-scale weighting-based edge-smoothing filter for single image dehazing,

S. K. Yadav and K. Sarawadekar, “Robust multi-scale weighting-based edge-smoothing filter for single image dehazing,”Pattern Recognit., vol. 149, p. 110 137, 2024

2024

[1] [1]

Videomage: Multi-subject and motion customization of text-to-video diffusion models,

C. Huang, Y . Wu, H. Chung, K. Chang, F. Yang, and Y . F. Wang, “Videomage: Multi-subject and motion customization of text-to-video diffusion models,” in CVPR, 2025

2025

[2] [2]

Cogvideox: Text-to-video diffusion models with an expert transformer,

Z. Yang et al., “Cogvideox: Text-to-video diffusion models with an expert transformer,” inICLR, 2025

2025

[3] [3]

Wonderworld: Interactive 3d scene generation from a single image,

H. Yu, H. Duan, C. Herrmann, W. T. Freeman, and J. Wu, “Wonderworld: Interactive 3d scene generation from a single image,” inCVPR, 2025

2025

[4] [4]

SVG: 3d stereoscopic video generation via denoising frame matrix,

P. Dai et al., “SVG: 3d stereoscopic video generation via denoising frame matrix,” inICLR, 2025

2025

[5] [5]

Particle guidance: Non-i.i.d. diverse sampling with diffusion models,

G. Corso, Y . Xu, V . D. Bortoli, R. Barzilay, and T. S. Jaakkola, “Particle guidance: Non-i.i.d. diverse sampling with diffusion models,” inICLR, 2024

2024

[6] [6]

Diverseflow: Sample- efficient diverse mode coverage in flows,

M. M. Morshed and V . Boddeti, “Diverseflow: Sample- efficient diverse mode coverage in flows,” inCVPR, 2025

2025

[7] [7]

Flow matching for generative modeling,

Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inICLR, 2023

2023

[8] [8]

Generative adversarial nets,

I. J. Goodfellow et al., “Generative adversarial nets,” in NIPS, 2014

2014

[9] [9]

Generat- ing videos with scene dynamics,

C. V ondrick, H. Pirsiavash, and A. Torralba, “Generat- ing videos with scene dynamics,” inNIPS, 2016

2016

[10] [10]

Density estimation using real NVP,

L. Dinh, J. Sohl-Dickstein, and S. Bengio, “Density estimation using real NVP,” inICLR, 2017

2017

[11] [11]

Videoflow: A conditional flow-based model for stochastic video generation,

M. Kumar et al., “Videoflow: A conditional flow-based model for stochastic video generation,” inICLR, 2020

2020

[12] [12]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inNeurIPS, 2020

2020

[13] [13]

Video diffusion models,

J. Ho, T. Salimans, A. A. Gritsenko, W. Chan, M. Norouzi, and D. J. Fleet, “Video diffusion models,” in NeurIPS, 2022. Data distribution 10 20 30 40 Noise distribution Steps Final latents Prediction & Decoding … … … Decoded videos Video latents … 50 (a) Extrapolation ofˆx1 =x t + (1−t)v θ(xt, t)from intermediate latent states and decoding them to video fr...

2022

[14] [14]

Flow straight and fast: Learning to generate and transfer data with rectified flow,

X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” inICLR, 2023

2023

[15] [15]

Wan: Open and advanced large-scale video generative models,

A. Wang et al., “Wan: Open and advanced large-scale video generative models,”CoRR, vol. abs/2503.20314, 2025

Pith/arXiv arXiv 2025

[16] [16]

Importance- weighted non-iid sampling for flow matching models,

X. Liu, R. B. Li, S. Wei, and T. Nguyen, “Importance- weighted non-iid sampling for flow matching models,” arXiv preprint arXiv:2511.17812, 2025

Pith/arXiv arXiv 2025

[17] [17]

Determinantal point pro- cesses for machine learning,

A. Kulesza, B. Taskar, et al., “Determinantal point pro- cesses for machine learning,”F oundations and Trends® in Machine Learning, vol. 5, no. 2–3, pp. 123–286, 2012

2012

[18] [18]

VideoPrism: A foundational visual encoder for video understanding,

L. Zhao et al., “VideoPrism: A foundational visual encoder for video understanding,” inICML, 2024

2024

[19] [19]

Learning transferable visual models from natural language supervision,

A. Radford et al., “Learning transferable visual models from natural language supervision,” inICML, 2021

2021

[20] [20]

The vendi score: A diversity evaluation metric for machine learning,

D. Friedman and A. B. Dieng, “The vendi score: A diversity evaluation metric for machine learning,”Trans. Mach. Learn. Res., 2023

2023

[21] [21]

EDEN: enhanced diffusion for high- quality large-motion video frame interpolation,

Z. Zhang et al., “EDEN: enhanced diffusion for high- quality large-motion video frame interpolation,” in CVPR, 2025

2025

[22] [22]

Natural color image enhancement and evaluation algorithm based on human visual system,

K. Huang, Q. Wang, and Z. Wu, “Natural color image enhancement and evaluation algorithm based on human visual system,”Comput. Vis. Image Underst., vol. 103, no. 1, pp. 52–63, 2006

2006

[23] [23]

Robust multi-scale weighting-based edge-smoothing filter for single image dehazing,

S. K. Yadav and K. Sarawadekar, “Robust multi-scale weighting-based edge-smoothing filter for single image dehazing,”Pattern Recognit., vol. 149, p. 110 137, 2024

2024