Pith. sign in

REVIEW 3 major objections 5 minor 23 references

Autoregressive video distillation works when the student already covers the teacher’s modes—and stays covering them while DMD sharpens.

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

2026-07-30 20:20 UTC pith:GHJH5PNW

load-bearing objection Solid empirical methods paper: match init to the DMD teacher and keep a CD anchor; the weak-teacher headline slightly overclaims what Table 4 isolates. the 3 major comments →

arxiv 2607.26811 v1 pith:GHJH5PNW submitted 2026-07-29 cs.CV

DistillAlign: Coordinating Mode Covering and Mode Seeking in Autoregressive Video Distillation

Jiaxing Li , Kai Zou , Cindy Zhou , Kaichen Huang , Junyao Gao , Zile Wang , Yang Liu , Bin Liu
show 2 more authors
Bo An Yangguang Li
This is my paper
classification cs.CV
keywords autoregressive video distillationdistribution matching distillationconsistency distillationmode coveringmode seekingreverse KLprecision and coveragefew-step video generation
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.

Standard pipelines first initialize a few-step causal video student, then refine it with distribution matching distillation (DMD). Those stages often chase different target distributions, and people pick the “best” initializer by visual scores such as VBench. This paper argues that is the wrong criterion. Because DMD is mode-seeking (reverse KL), it can only sharpen modes the student already covers; mismatched or collapsed initializations leave gradients unsupported and waste teacher capacity. The authors measure that gap with a teacher-normalized precision/coverage protocol in a shared video feature space, show that pure DMD still drifts toward high-probability teacher regions late in training, and fix both problems by matching init and DMD targets and training DMD jointly with a consistency-distillation mode-covering term. Under that recipe, even a small 1.3B DMD teacher beats larger 14B-refined baselines on quality, coverage, and diversity.

Core claim

Final quality in autoregressive video distillation is governed by distributional alignment, not by intermediate visual scores or teacher size alone: the pre-DMD student must cover the DMD teacher’s modes, and DMD must be paired with a mode-covering constraint so reverse-KL refinement does not collapse coverage and diversity. When both hold, a weaker matched teacher can outperform stronger mismatched pipelines.

What carries the argument

Joint distillation: L_joint = L_DMD + λ L_CD, plus a teacher-normalized precision/coverage protocol (re-noise initializer latents with a shared teacher, then k-NN support in V-JEPA2 feature space) that ranks initializers by mode coverage rather than raw sharpness.

Load-bearing premise

The claim rests on treating precision and coverage in one frozen video feature space, after a shared re-noising step, as a faithful stand-in for the semantic modes DMD can actually refine.

What would settle it

Hold the matched joint recipe fixed and swap or ablate the feature space and re-noising protocol: if initializer rankings and the link from higher coverage to better post-DMD VBench/diversity reverse or disappear under another encoder, radius rule, or no re-noising, the distributional story fails.

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

If this is right

  • Initialization checkpoints should be selected by matched teacher coverage, not peak VBench alone.
  • Scaling the DMD teacher helps less if the student never covered that teacher’s modes.
  • Pure reverse-KL DMD needs an explicit mode-covering anchor to avoid late diversity and quality drop.
  • Matched weak teachers can beat mismatched strong-teacher baselines in few-step causal video generation.

Where Pith is reading between the lines

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

  • The same cover-then-seek logic likely applies to other multi-stage score-distillation stacks beyond video (image, audio, 3D) wherever reverse-KL refinement follows a trajectory or consistency init.
  • If coverage is the bottleneck, collecting more diverse teacher samples for the init stage may beat simply swapping in a larger DMD teacher.
  • Online monitoring of coverage during DMD could serve as an early-stopping signal before visual scores peak and then fall.

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 argues that multi-stage autoregressive video distillation fails when pre-DMD initialization and DMD refinement target misaligned distributions, and when pure reverse-KL DMD is run long enough to collapse coverage. It introduces a teacher-normalized precision/coverage protocol in V-JEPA2 feature space to diagnose mode coverage independently of raw visual scores, shows via controlled teacher-source swaps and pipeline comparisons that matched mode-covering initializations (especially causal CD) support better DMD refinement, and proposes joint distillation L_DMD + λ L_CD to retain coverage while sharpening modes. Empirically, matched joint training improves VBench, coverage, and Vendi diversity; a full Wan-14B setting reaches the best scores, and a weak Wan-1.3B setting is reported to beat prior Wan-14B-refined baselines.

Significance. If the distributional account holds, the work reframes AR video distillation practice away from chasing intermediate VBench and toward explicit target matching and mode-cover/mode-seek coordination—an actionable design principle for few-step causal generators. Strengths include a clear two-hypothesis structure, controlled target-swap experiments (Tables 1 and 3), multi-pipeline coverage diagnostics (Figure 3, Table 7), training-dynamics evidence of late DMD drift versus joint anchoring (Figure 5, Tables 5–6 and weak-setting Tables 8–9), and a simple joint objective that is easy to adopt. The contribution is primarily empirical and systems-level rather than theoretical, but it is timely for real-time world-model and interactive video pipelines where both fidelity and diversity matter.

major comments (3)
  1. [Abstract; §4.2 Table 4; §1 Table 1] Abstract and §4.2 / Table 4 over-attribute the headline “Wan-1.3B DMD teacher outperforms Wan-14B-refined baselines.” Ours (weak) couples matched 1.3B data, causal CD init, and joint DMD+CD; Self-Forcing and Causal-Forcing are different recipes with 14B teachers and no joint anchor. Table 1 already shows matched pure DMD 1.3B→1.3B at 84.50, nearly identical to Ours (weak) 84.54, so most of the lift over mismatched Causal Forcing (83.89) is target matching, not joint loss or “alignment rivals 14B scale.” Please either (i) add a same-recipe ablation (matched CD init ± joint loss under both 1.3B and 14B teachers, reported side-by-side) or (ii) rewrite the claim as “full DistillAlign pipeline with a 1.3B teacher beats prior published 14B pipelines,” without implying a pure teacher-scale swap.
  2. [§4.1 Evaluation; §4.2 Table 4] In Table 4, precision/coverage for Ours (weak) are computed against the Wan-1.3B teacher while all 14B baselines and Ours (full) are scored against Wan-14B. §4.1 correctly warns that cross-teacher values are not on one absolute scale, yet the text still ranks “strongest coverage among distilled autoregressive models” in a way that invites direct comparison of 0.59 (vs 1.3B) to 0.29/0.53 (vs 14B). Restrict cross-method coverage ranking to rows that share the same reference teacher (e.g., emphasize Ours full 0.69 vs Causal-Forcing 0.29 / Self-Forcing 0.53), and report weak-setting coverage only within the 1.3B reference (as in Tables 8–9).
  3. [§3.1 Eqs. 8–11; Appendix A; Table 7] The causal story that higher teacher-normalized coverage causes better DMD refinement rests on the V-JEPA2 k-NN protocol (Sec. 3.1, Eqs. 8–11; Appendix A: ρ=0.9, shared 14B re-noise, k=5, N=M=256, 8-frame short window). Figure 8 and Table 7 make a good case that re-noising removes low-level confounds for ODE vs Causal DMD, but there is no sensitivity check to encoder (e.g., alternative video SSL or CLIP-style features), k, ρ, or temporal window. A compact ablation showing that initializer rankings and the CD-vs-DMD coverage gap are stable under at least one alternate feature space or (k, ρ) setting would substantially strengthen the load-bearing diagnostic.
minor comments (5)
  1. [Figure 1; Figure 2] Figure 1/2 labels contain typos (“Hyposis”, “Matche d”, “probabil ity”) and should be cleaned for camera-ready.
  2. [§2.3; §4.1] Eq. (6)–(7) present DMD as reverse KL on perturbed marginals; a brief note on how the online fake score is trained in the AR/causal setting (history conditioning, which frames enter the critic) would help reproducibility.
  3. [§3.3.2; §4.1 Implementation details] Joint loss uses the Stage-2 causal CD teacher while DMD uses the matched data-source teacher; state explicitly whether these coincide in the full 14B setting and how λ interacts when they differ.
  4. [§4.1; Tables 4–7] Diversity is raw Vendi on the same features as coverage; consider also reporting a teacher-relative diversity ratio so absolute scale is easier to interpret across tables.
  5. [§5.2] Related work §5.2 cites several concurrent mode-cover/mode-seek hybrids (f-Distill, rCM, ADM/DMDX, SGMD, HiAR, BiWM); a short explicit contrast of what is specific to causal multi-stage AR video (matched init across stages + CD anchor during DMD) would sharpen novelty.

Circularity Check

0 steps flagged

No significant circularity: empirical distillation methods paper with external benchmarks, not a self-sealing derivation.

full rationale

DistillAlign does not present a first-principles derivation whose outputs reduce to its inputs by construction. The central claims are empirical: (i) matching pre-DMD initialization support to the DMD teacher improves refinement, supported by a controlled teacher-source swap (Table 1) and pipeline comparison (Fig. 3, Table 7) that report external VBench scores; (ii) pure reverse-KL DMD can drift late even under matched targets (Table 2/6); (iii) joint loss L_joint = L_DMD + λ L_CD with ablated λ (Tables 5–6, 8–9) improves the quality–coverage–diversity trade-off. Precision/coverage (Eqs. 10–11) are post-hoc diagnostics in a frozen V-JEPA2 space, not training objectives optimized to force the headline. Mode-seeking vs mode-covering is standard reverse-KL vs consistency/ODE behavior from the cited literature, not a uniqueness theorem imported from the authors. Baselines (Self-Forcing, Causal-Forcing, CausVid) and VBench are external. No fitted parameter is renamed a prediction; no load-bearing self-citation chain closes the argument. Attribution debates about Table 4 (recipe vs teacher scale) are causal-identification issues, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central claim rests on standard generative-modeling assumptions about reverse-KL vs trajectory/consistency objectives, on Wan/VBench/V-JEPA2 tooling, and on several hand-chosen protocol and loss hyperparameters. No new physical entities are postulated; the main invented construct is the evaluation/normalization protocol used to operationalize ‘mode coverage.’

free parameters (5)
  • joint loss weight λ = 0.01
    Balances DMD and CD in L_joint; chosen by ablation and set to 0.01 as default because it gave the preferred quality/coverage trade-off.
  • teacher-normalized re-noise level ρ = 0.9
    Controls how much initializer outputs are re-noised before shared teacher refinement; fixed by authors for the coverage protocol.
  • k-NN support parameter k = 5
    Sets local teacher neighborhood radius for precision/coverage; fixed globally at 5.
  • distributional sample budget N=M = 256
    Number of student/teacher videos used to estimate precision/coverage; fixed at 256 on a 16×16 prompt-seed grid.
  • stage training lengths and data size = 5K / 2.5K / 1.5K; 25K samples
    AR diffusion 5K steps, CD 2.5K, joint 1.5K (ablations to 2.5K), 25K distillation samples per teacher source; procedural choices that affect reported optima.
axioms (5)
  • domain assumption DMD’s reverse-KL / score-difference objective is inherently mode-seeking and tends to reduce coverage under prolonged training.
    Taken from DMD literature and used as Hypothesis II (Sec. 2.3, Fig. 2, Tab. 2); not re-proved here beyond empirical drift plots.
  • domain assumption ODE/consistency-style regression onto teacher trajectories is mode-covering and preserves broader support at the cost of sharpness.
    Stated in Sec. 2.2 and used to justify CD as the coverage anchor in joint distillation.
  • ad hoc to paper Frozen V-JEPA2 features on a short temporal window represent the high-level semantic modes relevant to distillation success.
    Operational choice in Sec. 3.1 and Appendix A; the distributional protocol stands or falls with this representation.
  • ad hoc to paper Shared teacher re-noising mostly removes low-level fidelity confounds while preserving initializer-induced semantic support.
    Core justification for the teacher-normalized protocol (Sec. 3.1, Fig. 8, Sec. 4.3.3).
  • standard math Flow-matching / PF-ODE teacher maps and standard reverse-KL DMD gradient estimators are valid training targets for the student.
    Background Sec. 2 relies on established flow-matching and DMD formulations from cited work.
invented entities (2)
  • Teacher-normalized distributional evaluation protocol (precision/coverage after shared re-noising) no independent evidence
    purpose: Rank initializers by matched mode coverage rather than raw visual scores before DMD.
    Not a physical entity, but a new measurement construct central to the paper’s causal story; evidence for its validity is internal (correlation with post-DMD outcomes), not an external benchmark standard.
  • Joint DMD+CD distillation objective L_DMD + λ L_CD for AR video students no independent evidence
    purpose: Sharpen high-probability modes while anchoring teacher coverage to limit late DMD drift.
    A training recipe combining known losses; novelty is the coordinated use in this pipeline, not a new mathematical object with outside falsifiable mass/charge-like predictions.

pith-pipeline@v1.2.0-daily-grok45 · 23322 in / 3995 out tokens · 78758 ms · 2026-07-30T20:20:48.725859+00:00 · methodology

0 comments
read the original abstract

Existing autoregressive video distillation methods commonly adopt a Distribution Matching Distillation (DMD)-based multi-stage pipeline. However, they typically decouple the initialization and DMD stages -- which then pursue different target distributions -- and judge the intermediate student mainly by visual scores such as VBench. In this paper, we revisit this design from a distributional perspective. Given the mode-seeking nature of the distribution matching loss, a good initialization should match the mode coverage of the target DMD teacher, rather than merely pursuing high quality. To analyze this, we introduce a distributional evaluation protocol that measures precision and coverage between student and teacher distributions in a shared latent space. It exposes differences hidden by visual scores: some initializations reach high precision but low coverage, leading to suboptimal refinement, while mode-covering ones preserve broader support. Furthermore, even when the target distributions are aligned, DMD's reverse-KL objective can still drive the student toward high-probability teacher regions in late training, reducing coverage and diversity. To address this, we propose joint distillation, which combines DMD's mode-seeking objective with a Consistency Distillation-based mode-covering constraint. Experiments show that our method improves generation quality, coverage, and diversity; notably, even with a Wan-1.3B DMD teacher, it outperforms baselines refined with Wan-14B, underscoring the importance of distributional alignment in autoregressive video distillation.

Figures

Figures reproduced from arXiv: 2607.26811 by Bin Liu, Bo An, Cindy Zhou, Jiaxing Li, Junyao Gao, Kaichen Huang, Kai Zou, Yangguang Li, Yang Liu, Zile Wang.

Figure 1
Figure 1. Figure 1: Illustration of Hypothesis I. Mismatched target distributions across stages can lead to suboptimal refinement [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of Hypothesis II. Even with matched initialization, a long-term reverse-KL objective can push the student distribution toward high-probability regions of the teacher [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual diversity comparison across different distillation pipelines. For each pipeline, we sample 16 fixed seeds and visualize the middle frame of each generated 81-frame video as a grid. 3.2.2 RETHINKING DISTILLATION PIPES FROM A DISTRIBUTIONAL PERSPECTIVE Bidirectional Model AR Diffusion Distribution Matching Distillation Bidirectional Model ODE Distillation Distribution Matching Distillation Student Ini… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison across different distillation pipelines. All pipelines use Wan-14B as the DMD teacher and use its sampled distillation data as the initialization target. When the data source and target distribution are already aligned, how do different initialization methods affect coverage and subsequent DMD refinement? As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Distribution evolution under mode-covering and mode-seeking objectives. We visual￾ize checkpoint distributions in a shared V-JEPA2 PCA space and track precision and coverage over training. Causal CD expands teacher coverage, pure DMD contracts toward high-density teacher regions, and joint distillation preserves coverage while refining samples. The left panel shows the evolution of the Causal CD initialize… view at source ↗
Figure 6
Figure 6. Figure 6: provides qualitative examples consistent with the quantitative trend. Compared with prior distilled autoregressive baselines, our model better preserves object identity and motion details across frames, while avoiding the overly smoothed appearance typical of mode-covering initial￾ization alone. Wan2.1-1.3B Self Forcing Causal Forcing Ours (1.3B DMD Teacher) Ours (14B DMD Teacher) Wan2.1-1.3B Self Forcing … view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of sample diversity. Under fixed prompts and different seeds, our method preserves richer appearance and motion variations while maintaining high visual fidelity [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Effect of teacher-normalized re-noising. We compare raw initializer samples before and after ap￾plying a shared teacher refinement. Re-noising preserves the initializer-induced scene semantics while reducing nuisance variation in low-level fidelity, making the distributional comparison less sensitive to raw sharpness dif￾ferences. 10 [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] 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

23 extracted references · 16 linked inside Pith

  1. [1]

    V-JEPA 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985,

    Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Mojtaba Komeili, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, et al. V-JEPA 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985,

    Pith/arXiv arXiv

  2. [5]

    Dynamics- boost: Dynamic plausible video generation via annotation-free continuation preference optimiza- tion

    12 arXivPreprint, 2026 Jiaxing Li, Jiepeng Wang, Junyao Gao, Yang Liu, Eric Li, Bo An, and Hao-Xiang Guo. Dynamics- boost: Dynamic plausible video generation via annotation-free continuation preference optimiza- tion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20024–20033,

    2026

  3. [8]

    Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003,

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003,

    Pith/arXiv arXiv

  4. [9]

    Knowledge distillation in iterative generative models for improved sampling speed.arXiv preprint arXiv:2101.02388,

    Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed.arXiv preprint arXiv:2101.02388,

    Pith/arXiv arXiv

  5. [10]

    Riemann Dynamics

    Accessed: 2026-06-20. Riemann Dynamics. Matrix-game 3.5: Enhancing real-time streaming interactive world models with patch memory. Project page,

    2026

  6. [11]

    Shaohao Rui, Xiaofeng Mao, Zhanyu Zhang, Peijia Lin, Yansong Zhu, Yibo Zhang, Haibin Wan, and Weijie Ma

    URLhttps://matrix-game-v3-5.github.io/. Shaohao Rui, Xiaofeng Mao, Zhanyu Zhang, Peijia Lin, Yansong Zhu, Yibo Zhang, Haibin Wan, and Weijie Ma. Biwm: Advancing open-source interactive video world models with bidirectional autoregression.arXiv preprint arXiv:2606.10135,

    Pith/arXiv arXiv

  7. [12]

    Advancing open-source world models.arXiv preprint arXiv:2601.20540,

    Robbyant Team, Zelin Gao, Qiuyu Wang, Yanhong Zeng, Jiapeng Zhu, Ka Leong Cheng, Yixuan Li, Hanlin Wang, Yinghao Xu, Shuailei Ma, et al. Advancing open-source world models.arXiv preprint arXiv:2601.20540,

    Pith/arXiv arXiv

  8. [13]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,

    Wan Team. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,

    Pith/arXiv arXiv

  9. [14]

    Videolcm: Video latent consistency model.arXiv preprint arXiv:2312.09109,

    Xiang Wang, Shiwei Zhang, Han Zhang, Yu Liu, Yingya Zhang, Changxin Gao, and Nong Sang. Videolcm: Video latent consistency model.arXiv preprint arXiv:2312.09109,

    Pith/arXiv arXiv

  10. [15]

    Matrix-game 3.0: Real-time and streaming interactive world model with long-horizon memory.arXiv preprint arXiv:2604.08995,

    13 arXivPreprint, 2026 Zile Wang, Zexiang Liu, Jiaxing Li, Kaichen Huang, Baixin Xu, Fei Kang, Mengyin An, Peiyu Wang, Biao Jiang, Yichen Wei, et al. Matrix-game 3.0: Real-time and streaming interactive world model with long-horizon memory.arXiv preprint arXiv:2604.08995,

    Pith/arXiv arXiv 2026

  11. [16]

    One-step diffusion models withf-divergence distribution matching.arXiv preprint arXiv:2502.15681,

    Yilun Xu, Weili Nie, and Arash Vahdat. One-step diffusion models withf-divergence distribution matching.arXiv preprint arXiv:2502.15681,

    Pith/arXiv arXiv

  12. [17]

    Longlive: Real-time interactive long video generation

    Shuai Yang, Wei Huang, Ruihang Chu, Yicheng Xiao, Yuyang Zhao, Xianbang Wang, Muyang Li, Enze Xie, Yingcong Chen, Yao Lu, et al. Longlive: Real-time interactive long video generation. arXiv preprint arXiv:2509.22622,

    Pith/arXiv arXiv

  13. [18]

    Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072,

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072,

    Pith/arXiv arXiv

  14. [19]

    Causal forcing++: Scalable few-step autoregressive diffusion distil- lation for real-time interactive video generation.arXiv preprint arXiv:2605.15141,

    Min Zhao, Hongzhou Zhu, Kaiwen Zheng, Zihan Zhou, Bokai Yan, Xinyuan Li, Xiao Yang, Chongxuan Li, and Jun Zhu. Causal forcing++: Scalable few-step autoregressive diffusion distil- lation for real-time interactive video generation.arXiv preprint arXiv:2605.15141,

    Pith/arXiv arXiv

  15. [20]

    Hiar: Efficient au- toregressive long video generation via hierarchical denoising.arXiv preprint arXiv:2603.08703,

    Kai Zou, Dian Zheng, Hongbo Liu, Tiankai Hang, Bin Liu, and Nenghai Yu. Hiar: Efficient au- toregressive long video generation via hierarchical denoising.arXiv preprint arXiv:2603.08703,

    arXiv

  16. [21]

    14 arXivPreprint, 2026 A MOREDETAILS OFCOVERAGEMETRICS This section specifies the complete evaluation pipeline used for the precision, coverage, and diversity results in the paper. We distinguish thenormalization teacherT norm, which removes nuisance differences among raw initializer outputs, from thereference teacherTref, whose sample distribution define...

    2026

  17. [22]

    on the same normalized V-JEPA2 descriptors. For a method feature matrixU∈R N×2560 , we form the shifted cosine kernel K= U U⊤ +11 ⊤ 2 , ¯K= K N .(18) If{λ q}are the positive eigenvalues of ¯K, the reported score is Vendi(U) = exp − X q λq logλ q ! .(19) We report this raw effective-rank value, not a ratio to the teacher. It is a within-method diversity st...

    2026

  18. [23]

    For fair comparison, all methods in the same figure use the same prompt, and when applicable, the same set of fixed random seeds

    VBench Metrics↑Distributional Metrics↑VBench Metrics↑Distributional Metrics↑ Total Quality Semantic Precision Coverage Diversity Total Quality Semantic Precision Coverage Diversity 0 82.16 83.38 77.29 0.61 0.52 1.301 82.16 83.38 77.29 0.61 0.52 1.301 500 83.84 85.07 78.92 0.66 0.52 1.295 84.44 85.52 80.09 0.68 0.48 1.293 1000 84.17 85.38 79.29 0.69 0.51 1...

    2026

  19. [2022]

    Rolling forcing: Autoregressive long video diffusion in real time.arXiv preprint arXiv:2509.25161,

    Kunhao Liu, Wenbo Hu, Jiale Xu, Ying Shan, and Shijian Lu. Rolling forcing: Autoregressive long video diffusion in real time.arXiv preprint arXiv:2509.25161,

    Pith/arXiv arXiv

  20. [2023]

    Charactershot: Controllable and consistent 4d character animation.arXiv preprint arXiv:2508.07409,

    Junyao Gao, Jiaxing Li, Wenran Liu, Yanhong Zeng, Fei Shen, Kai Chen, Yanan Sun, and Cairong Zhao. Charactershot: Controllable and consistent 4d character animation.arXiv preprint arXiv:2508.07409,

    Pith/arXiv arXiv

  21. [2024]

    Context forcing: Consistent autoregressive video generation with long context.arXiv preprint arXiv:2602.06028,

    Shuo Chen, Cong Wei, Sun Sun, Ping Nie, Kai Zhou, Ge Zhang, Ming-Hsuan Yang, and Wenhu Chen. Context forcing: Consistent autoregressive video generation with long context.arXiv preprint arXiv:2602.06028,

    arXiv

  22. [2025]

    Matrix-game 2.0: An open-source real-time and stream- ing interactive world model.arXiv preprint arXiv:2508.13009,

    Xianglong He, Chunli Peng, Zexiang Liu, Boyang Wang, Yifan Zhang, Qi Cui, Fei Kang, Biao Jiang, Mengyin An, Yangyang Ren, et al. Matrix-game 2.0: An open-source real-time and stream- ing interactive world model.arXiv preprint arXiv:2508.13009,

    Pith/arXiv arXiv

  23. [2026]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

    Pith/arXiv arXiv