Pith. sign in

REVIEW 5 major objections 6 minor 92 references

Video generation can run as a live, steerable stream: change the prompt mid-rollout and see it on screen in under a second, for hours, at 4K/24 FPS.

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 23:45 UTC pith:PRCDSH5Y

load-bearing objection Solid end-to-end real-time interactive video system with real Arena wins; the hour-scale stability claim is still mostly a single pretty trajectory. the 5 major comments →

arxiv 2607.26694 v1 pith:PRCDSH5Y submitted 2026-07-29 cs.CV

Visko Orbis 1.0: A Live Model for Real-Time Interactive Long Video Generation

Xiangbo Gao , Siyuan Yang , Ping He , Mingyang Wu , Yuheng Wu , Yushen Zuo , Jiongze Yu , Ryan Cui
show 7 more authors
Hongyuan Hua Devin Ma Xiao Jin Yubo Yuan Qing Yin Jie Yang Zhengzhong Tu
This is my paper
classification cs.CV
keywords live modelsreal-time video generationinteractive long videochunk-wise streamingbounded multi-scale memoryprompt switchingstreaming super-resolutioncausal video diffusion
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that video models should behave like live processes, not one-shot clip factories. Visko Orbis 1.0 generates long video in successive chunks while the user can rewrite the prompt at any moment; the new instruction lands on the next uncommitted chunk and shows up in the delivered stream in under a second on average. A bounded multi-scale memory keeps subjects, scenes, and style coherent across chunks so rollouts can run for an hour without obvious quality or color drift. A distilled chunk-wise generator plus a streaming super-resolution path, served on an optimized multi-GPU stack, produces native 832×480 structure and delivers 4K at 24 FPS in real time. On multi-event one- to three-minute benchmarks and a long-form human Arena, the system leads compared real-time interactive systems on overall preference and temporal stability, and tops several automated quality scores.

Core claim

The authors claim that a single causal chunk-wise latent-flow Live Model—with carried visual state, event-aligned training, few-step distillation and reward alignment, and streaming upscaling—can support interactive long-form T2V, I2V, and continuation with mid-generation prompt switches, real-time 4K/24 FPS delivery, and hour-scale consistency without restarting the rollout.

What carries the argument

Bounded multi-scale memory inside a persistent chunk-wise latent rollout: recent history stays high-resolution, older spans compress under a fixed token budget, and evicted detail consolidates into a fixed-capacity learned state, so per-chunk cost stays bounded while subjects, scenes, and style carry forward.

Load-bearing premise

The claim that this memory and drift control truly hold for hour-long diverse video rests mainly on a single long diagnostic rollout, while the controlled head-to-head numbers cover only about one- to three-minute multi-event clips.

What would settle it

Run many multi-seed, multi-prompt hour-scale interactive sessions across varied content, log DOVER, alignment, saturation, and human temporal-stability over the full hour, and check whether quality and color stay stable and whether prompt switches remain visible sub-second without restarts—if drift or failure rates climb systematically after a few minutes, the long-horizon Live Model claim fails.

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

If this is right

  • Interactive storytelling, live streaming, games, and virtual companions can treat video as a continuous steerable feed rather than a finished clip to regenerate.
  • T2V, I2V, and video continuation can share one causal continuation interface instead of separate offline pipelines.
  • Serving co-design (state reuse, sequence-parallel batch-one inference, progressive decode and 4K upscale) becomes part of the model contract, not an afterthought.
  • Evaluation shifts from short isolated clips to complete multi-event streams scored for continuity across prompt updates and long-horizon drift.

Where Pith is reading between the lines

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

  • If hour-scale stability generalizes, offline “regenerate the whole clip” UX may give way to director-style live control as the default creative loop.
  • The same bounded-memory plus chunk-boundary conditioning pattern could transfer to other continuous media (audio worlds, embodied simulators) that must accept mid-stream intent changes.
  • A standardized multi-hour interactive stress suite with uncertainty bands would become the decisive comparison layer for any system claiming Live Model status.

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

5 major / 6 minor

Summary. Visko Orbis 1.0 is presented as a Live Model for real-time interactive long-video generation: a chunk-wise causal latent-flow generator with bounded multi-scale memory, event-aligned multi-stage training, few-step distillation plus GRPO alignment, content-adaptive drift controls, and a streaming 4K super-resolution path. The system supports T2V/I2V/V2V continuation, multilingual prompts, and in-generation prompt switching with claimed sub-second visible response, delivering 4K at 24 FPS. Empirically, on a 74-case multi-event benchmark of roughly 1–3 minutes, it leads several prompt-independent metrics (DOVER aesthetic/technical; VideoAlign visual/motion) and, in a long-form human Arena on 60–120 s clips, obtains the highest overall-preference and temporal-stability Elo point estimates among the compared real-time interactive systems.

Significance. Real-time, steerable, long-horizon video generation is a genuine systems frontier beyond offline short-clip models, and the manuscript assembles a full stack—data governance, streaming adaptation, memory, distillation/RL, serving, and progressive SR—under a coherent Live Model contract. If the interactive latency, short-form quality leadership, and long-horizon stability claims hold under stronger protocols, the work would be a useful reference system for interactive video. Strengths include a clear causal rollout formulation, an explicit multi-stage training curriculum, and external automated plus human Arena evaluation rather than purely internal scores. The contribution is primarily systems integration and empirical demonstration rather than a single new theoretical mechanism; its lasting value depends on how rigorously the long-horizon and interaction claims are substantiated.

major comments (5)
  1. [Abstract; §1; Figure 1; §5] Abstract, §1, and Figure 1 claim hour-scale rollouts “without evident quality or color drift” via bounded multi-scale memory. The only controlled quantitative protocol (§5) uses ~1–3 minute multi-event videos (74 cases); the Arena (Table 2) uses 60–120 s clips. Hour-scale support is essentially one diagnostic trajectory’s DOVER/VideoAlign/HSV curves in Fig. 1, with no multi-seed, multi-prompt, multi-content protocol, uncertainty, or failure-case analysis. Either provide a repeated long-horizon evaluation with uncertainty, or substantially qualify the hour-scale claim so it does not outrun the controlled evidence.
  2. [§3 Bounded Multi-Scale Memory; Long-Horizon Reliability; Distillation and RL; §4 Content-Adaptive Drift Stabilization; §] §3 lists bounded multi-scale memory, history perturbation / rollout-calibrated augmentation, GRPO reward mixture, latent world-model physics alignment, and §4 content-adaptive drift profiles as load-bearing ingredients, yet §5 reports only end-to-end system scores with no ablations isolating memory vs. inference-time drift controls, distillation stages, or reward terms. Without these, it is unclear which components drive temporal-stability leadership (Table 2) versus short-clip quality (Table 1). At minimum, ablate memory capacity/compression and the drift-control profiles on the multi-event benchmark and on extended rollouts.
  3. [§5 Table 1 Panel B] Table 1 Panel B is explicitly labeled “descriptive rather than a formal cross-system rank because prompt suites differ,” yet the text still highlights Orbis’s largest raw VideoAlign text-alignment, overall reward, and HeliosBench semantic scores. If prompts are unmatched, these cannot support comparative leadership claims. Either harmonize prompts across systems and re-rank, or move Panel B out of comparative framing and restrict claims to Panel A and the human Arena.
  4. [§5 Table 2; Figure 4] The human Arena (Table 2, Figure 4) is central to the abstract’s SOTA claim, but the baseline set differs from Table 1 (e.g., HappyOyster, Odyssey, PixVerse-R1, LongLive-2 appear only in the Arena; several Table 1 systems are absent). Rater count, pairing design, clip sampling, and confidence/intervals around Elo point estimates are not reported. Please document the Arena protocol fully and, where possible, align quantitative and human baseline sets so overall-preference/temporal-stability leadership is interpretable.
  5. [§1; §4; Conclusion] Interactive latency (“update becomes visible in under one second on average,” §1/Conclusion) and real-time 4K@24 FPS are core Live Model claims, but §4–5 give no measurement protocol: hardware topology, batching, time-to-first-visible-frame after a mid-rollout prompt switch, jitter, or end-to-end pipeline breakdown (generation vs. decode vs. SR vs. encode). Report these measurements with the serving configuration used for the quality evaluations; otherwise the real-time interaction claim remains anecdotal relative to the metric tables.
minor comments (6)
  1. [§3 Model Formulation; Eq. (1)] Equation (1) defines the rectified-flow loss, but the notation q for conditioning is overloaded across bidirectional (q=c) and streaming (q=(c_k,r_k,H_k)) regimes without a single explicit conditional density statement tying p_θ(z_k|H_k,c_k,r_k) to the training objective.
  2. [§4 Physics-Aware Inference Alignment; §5] Physics-aware inference alignment via a latent world-model reward (§4, citing WMReward) is listed as a contribution but never appears in §5 metrics or ablations; either evaluate its effect or demote it in the contribution list.
  3. [Figure 1] Figure 1’s hour-long qualitative strip and metric trajectories are hard to read in grayscale; consider separating aesthetic, instruction-following, and saturation panels and stating which comparison systems the lighter curves denote.
  4. [§1; §5] Several related real-time/long-video systems are cited in §1 but omitted from both tables; a short “systems not compared” paragraph with reasons (API access, non-real-time, different conditioning) would help.
  5. [Abstract] Minor consistency: abstract says “real-time 4K video generation at 24 FPS” and “highest overall-preference and temporal-stability,” which matches Table 2, but the abstract’s DOVER/VideoAlign leadership wording in the full abstract block should stay aligned with the Table 1 qualification on prompt-aware scores.
  6. [§2] Typos/style: “borderlineautomaticdecisionsandprovidescalibratedpreferencejudgments” and similar spacing errors in §2 Human-Preference Curation; clean PDF text extraction issues before camera-ready.

Circularity Check

1 steps flagged

Empirical systems paper; no derivation-by-construction. Only mild non-load-bearing self-framing via the authors’ Live Models blog.

specific steps
  1. self citation load bearing [§1 Introduction, Live Models framing; Ref. [24]]
    "These requirements reflect a broader shift that we have articulated as Live Models[24]: foundation models that execute as persistent processes rather than answering bounded queries. A Live Model generates in real time, under the clock of a world in motion; interacts continuously, folding new input into a process already underway; and maintains an ongoing existence, carrying state forward as long as it runs."

    The paper’s central product category (‘Live Model’) is justified by a citation whose authors overlap the present team (Tu, Yang, Yin). This is framing/terminology, not a uniqueness theorem or a fitted quantity renamed as prediction: Arena Elo, DOVER, and VideoAlign results do not reduce to [24]. Mild and non-load-bearing; does not force the empirical claims.

full rationale

Visko Orbis 1.0 is a systems/ML paper whose load-bearing claims are empirical (real-time 4K@24FPS interactive long video, Arena Elo leadership, DOVER/VideoAlign scores, hour-scale diagnostic curves). There is no first-principles derivation chain in which a fitted parameter or self-defined quantity is relabeled as an independent prediction. Training objectives (rectified flow, CFG distillation, consistency/DMD, GRPO) and architectural choices (chunk-wise causal generation, bounded multi-scale memory, streaming SR) are standard constructions evaluated against external instruments (DOVER, VideoAlign, HPSv3, HeliosBench) and a human Arena with named baselines. The sole mild circularity-adjacent element is terminological self-framing: the ‘Live Model’ contract is introduced by citing the authors’ own blog [24], which defines the paradigm the system is said to realize. That citation does not force the metric or Arena results, does not import a uniqueness theorem, and does not make any reported score true by construction. Weaknesses of the hour-scale evidence (single diagnostic trajectory vs. 1–3 min controlled protocol) are generalization/evaluation gaps, not circularity. Score 1 reflects one non-load-bearing self-citation for framing only.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 3 invented entities

Load-bearing content is engineering assumptions and undisclosed fitted choices, not mathematical axioms. The central claims rest on: causal chunk factorization of video; adequacy of bounded multi-scale memory for long-horizon identity; validity of automatic and Arena metrics as proxies for live quality; and an unreproducible proprietary data/serving stack. Free parameters are numerous but mostly unreported (memory budgets, distillation steps, reward weights, drift-control profiles).

free parameters (6)
  • Memory token budget and multi-scale compression schedule
    Bounded multi-scale memory keeps recent latents fine and older spans compressed under a fixed token budget; the budget and compression policy are central to the no-drift claim but not numerically specified.
  • Few-step sampler step count and CFG distillation scale
    Real-time claims depend on a distilled conditional-only few-step student; step count and embedded guidance scale are chosen for speed/quality tradeoff and not reported as a fixed public recipe.
  • GRPO reward mixture weights (visual, motion, text-video alignment)
    Post-training combines normalized rewards into a relative advantage; weights/normalization affect reported quality leadership.
  • Content-adaptive drift-control profiles
    Inference uses content-specific combinations of history perturbation, RoPE scaling, attention reweighting, VAE refresh, latent-stat correction, and negative-prompt routing; profiles are hand-designed stabilizers for long rollouts.
  • Native resolution and chunk duration (832×480 generator; event lengths 10/20/30s) = 832×480 native; 4K delivery; 24 FPS; events 10–30s
    System and eval geometry are design choices that condition FPS, memory, and benchmark scores.
  • Data filtering thresholds and rebalancing ledger weights
    Multi-stage safety/quality filters and stage-specific rebalancing determine the training distribution; thresholds are calibrated by content type and not released.
axioms (5)
  • domain assumption Video generation can be factored as a causal product of chunk conditionals p(z_k | H_k, c_k, r_k) with deterministic history updates, so prompt updates need only affect uncommitted chunks.
    Stated in Model Formulation; standard for autoregressive/causal video systems but load-bearing for interactive control without restart.
  • domain assumption A fixed-capacity multi-scale memory plus occasional consolidation preserves subject/scene/style identity over hour-scale rollouts without unbounded compute.
    Core long-horizon claim in abstract/intro and Bounded Multi-Scale Memory section; only weakly stress-tested beyond minutes.
  • domain assumption Automatic metrics (DOVER, VideoAlign, HPSv3, HeliosBench) and pairwise Arena Elo on 60–120s clips are adequate proxies for live interactive long-video quality.
    Evaluation protocol Section 5; authors partly caveat prompt-suite mismatch but still headline leadership.
  • domain assumption Rectified-flow training, CFG distillation, consistency distillation, self-forcing DMD, and GRPO-style RL transfer from cited short/causal settings to this live stack.
    Post-training section relies on prior methods [8–11,40–48] without new theoretical guarantees.
  • ad hoc to paper Super-resolution should restore detail without inventing semantic state absent from native generation.
    Explicit design stance in Video Super-Resolution; evaluation separates native vs delivered-4K evidence, but user-facing claims emphasize 4K output.
invented entities (3)
  • Live Model (as a product/system contract for persistent generative processes) no independent evidence
    purpose: Frames the system boundary at delivered video with continuous interaction and ongoing state rather than offline clip queries.
    Introduced via authors’ blog [24] and used throughout; mostly a framing layer over streaming causal generation, not an independently measured scientific object.
  • Bounded multi-scale memory M_k with eviction-time consolidation no independent evidence
    purpose: Carry long-range visual state across chunks under fixed compute/memory.
    Described as reading persistent memory as an extra context tier and writing only on eviction; related to cited memory papers but specific Orbis constructor is new and not externally validated.
  • Latent world-model reward for physics-aware inference alignment no independent evidence
    purpose: Score candidate futures for predictive consistency to steer motion/object dynamics at inference.
    Invokes WMReward principle [75]; little architecture, training, or ablation detail in this manuscript, so it functions as an added postulated module.

pith-pipeline@v1.2.0-daily-grok45 · 21896 in / 4330 out tokens · 89042 ms · 2026-07-30T23:45:59.586807+00:00 · methodology

0 comments
read the original abstract

We present Visko Orbis 1.0, a Live Model for real-time, interactive long-video generation. Users can change the prompt at any moment during generation, and the update becomes visible in real time. Visko Orbis 1.0 supports long-form text-to-video, image-to-video, and video continuation, with multilingual prompts and prompt switching while generation is in progress. A bounded multi-scale memory preserves subjects, scenes, and style across chunks, sustaining hour-scale rollouts without evident quality or color drift. Built on a distilled chunk-wise streaming generator and a streaming video upscaler, Visko Orbis 1.0 delivers real-time 4K video generation at 24 FPS using an optimized GPU serving engine. In long-form Arena comparisons, Visko Orbis 1.0 obtains the highest overall-preference and temporal-stability ratings among state-of-the-art real-time interactive video-generation systems.

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

92 extracted references · 43 linked inside Pith

  1. [1]

    Video generation models as world simulators,

    T. Brooks, B. Peebles, C. Holmeset al., “Video generation models as world simulators,” OpenAI Technical Report, https://openai.com/index/video-generation-models-as-world-simulators/, 2024

    2024

  2. [2]

    Veo: Our leading video generation model,

    Google DeepMind, “Veo: Our leading video generation model,” https://deepmind.google/models/veo/, 2024, accessed July 15, 2026

    2024

  3. [3]

    Seedance 1.0: Exploring the boundaries of video generation models,

    Y. Gaoet al., “Seedance 1.0: Exploring the boundaries of video generation models,”arXiv preprint arXiv:2506.09113, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Kling-omni technical report,

    Kling Team, “Kling-omni technical report,”arXiv preprint arXiv:2512.16776, 2025

    Pith/arXiv arXiv 2025

  5. [5]

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

    Wan Team, “Wan: Open and advanced large-scale video generative models,”arXiv preprint arXiv:2503.20314, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Hunyuanvideo: A systematic framework for large video generative models,

    W. Konget al., “Hunyuanvideo: A systematic framework for large video generative models,”arXiv preprint arXiv:2412.03603, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Longcat-video technical report,

    Meituan LongCat Team, X. Cai, Q. Huang, Z. Kanget al., “Longcat-video technical report,”arXiv preprint arXiv:2510.22200, 2025

    arXiv 2025

  8. [8]

    From slow bidirectional to fast autoregressive video diffusion models,

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

    2025

  9. [9]

    Self forcing: Bridging the train-test gap in autoregressive video diffusion,

    X. Huang, Z. Li, G. He, M. Zhou, and E. Shechtman, “Self forcing: Bridging the train-test gap in autoregressive video diffusion,”arXiv preprint arXiv:2506.08009, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Rolling forcing: Autoregressive long video diffusion in real time,

    K. Liu, W. Hu, J. Xu, Y. Shan, and S. Lu, “Rolling forcing: Autoregressive long video diffusion in real time,”arXiv preprint arXiv:2509.25161, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Causal forcing: Autoregressive diffusion distillation done right for high-quality real-time interactive video generation,

    H. Zhu, M. Zhao, G. He, H. Su, C. Li, and J. Zhu, “Causal forcing: Autoregressive diffusion distillation done right for high-quality real-time interactive video generation,”arXiv preprint arXiv:2602.02214, 2026

    Pith/arXiv arXiv 2026

  12. [12]

    Oasis: A universe in a transformer,

    Decart, J. Quevedo, Q. McIntyre, S. Campbell, X. Chen, and R. Wachen, “Oasis: A universe in a transformer,” https: //oasis-model.github.io/, 2024

    2024

  13. [13]

    Longlive: Real-time interactive long video generation,

    S. Yanget al., “Longlive: Real-time interactive long video generation,”arXiv preprint arXiv:2509.22622, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Matrix-game 2.0: An open-source, real-time, and streaming interactive world model,

    X. Heet al., “Matrix-game 2.0: An open-source, real-time, and streaming interactive world model,”arXiv preprint arXiv:2508.13009, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    Krea realtime 14b: Real-time, long-form ai video generation,

    E. Millon, “Krea realtime 14b: Real-time, long-form ai video generation,” https://www.krea.ai/blog/krea-realtime-14b, 2025

    2025

  16. [16]

    Helios: Real real-time long video generation model,

    S. Yuan, Y. Yin, Z. Li, X. Huang, X. Yang, and L. Yuan, “Helios: Real real-time long video generation model,”arXiv preprint arXiv:2603.04379, 2026

    arXiv 2026

  17. [17]

    Vidu s1: A real-time interactive video generation model,

    J. Zhanget al., “Vidu s1: A real-time interactive video generation model,”arXiv preprint arXiv:2607.03118, 2026

    Pith/arXiv arXiv 2026

  18. [18]

    Frame context packing and drift prevention in next-frame-prediction video diffusion models,

    L. Zhang, S. Cai, M. Li, G. Wetzstein, and M. Agrawala, “Frame context packing and drift prevention in next-frame-prediction video diffusion models,”arXiv preprint arXiv:2504.12626, 2025

    arXiv 2025

  19. [19]

    Long-context autoregressive video modeling with next-frame prediction,

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

    Pith/arXiv arXiv 2025

  20. [20]

    Memflow: Flowing adaptive memory for consistent and efficient long video narratives,

    S. Ji, X. Chen, S. Yang, X. Tao, P. Wan, and H. Zhao, “Memflow: Flowing adaptive memory for consistent and efficient long video narratives,”arXiv preprint arXiv:2512.14699, 2025

    arXiv 2025

  21. [21]

    Videossm: Autoregressive long video generation with hybrid state-space memory,

    Y. Yu, X. Wu, X. Hu, T. Huet al., “Videossm: Autoregressive long video generation with hybrid state-space memory,” arXiv preprint arXiv:2512.04519, 2025

    arXiv 2025

  22. [22]

    Fademem: Distance-aware memory consolidation for autoregressive video diffusion,

    Y. Lu, J. Yang, P. Koniusz, Y. Song, and Y. Yang, “Fademem: Distance-aware memory consolidation for autoregressive video diffusion,”arXiv preprint arXiv:2606.10671, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    Echo-infinity: Learning evolving memory for real-time infinite video generation,

    Y. Bian, Z. Xue, S. Zhang, S. Zhanget al., “Echo-infinity: Learning evolving memory for real-time infinite video generation,” arXiv preprint arXiv:2606.04527, 2026

    Pith/arXiv arXiv 2026

  24. [24]

    Introducing live models: A frontier for world models,

    Z. Tu, J. Yang, and Q. Yin, “Introducing live models: A frontier for world models,” Visko Blog, https://www.visko.ai/blog/ introducing-live-models, 2026, july 2026

    2026

  25. [25]

    Internvid: A large-scale video-text dataset for multimodal understanding and generation,

    Y. Wang, Y. He, Y. Li, K. Liet al., “Internvid: A large-scale video-text dataset for multimodal understanding and generation,” arXiv preprint arXiv:2307.06942, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    Panda-70m: Captioning 70m videos with multiple cross-modality teachers,

    T.-S. Chen, A. Siarohin, W. Menapace, E. Deynekaet al., “Panda-70m: Captioning 70m videos with multiple cross-modality teachers,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    2024

  27. [27]

    Stable video diffusion: Scaling latent video diffusion models to large datasets,

    A. Blattmann, T. Dockhorn, S. Kulal, D. Mendelevitchet al., “Stable video diffusion: Scaling latent video diffusion models to large datasets,”arXiv preprint arXiv:2311.15127, 2023

    Pith/arXiv arXiv 2023

  28. [28]

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

    Z. Yang, J. Teng, W. Zheng, M. Dinget al., “Cogvideox: Text-to-video diffusion models with an expert transformer,” in International Conference on Learning Representations, 2025

    2025

  29. [29]

    Movie gen: A cast of media foundation models,

    A. Polyak, A. Zohar, A. Brown, A. Tjandraet al., “Movie gen: A cast of media foundation models,”arXiv preprint arXiv:2410.13720, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Datasheets for datasets,

    T. Gebru, J. Morgenstern, B. Vecchione, J. W. Vaughan, H. Wallach, H. Daumé III, and K. Crawford, “Datasheets for datasets,”arXiv preprint arXiv:1803.09010, 2018

    Pith/arXiv arXiv 2018

  31. [31]

    Data cards: Purposeful and transparent dataset documentation for responsible ai,

    M. Pushkarna, A. Zaldivar, and O. Kjartansson, “Data cards: Purposeful and transparent dataset documentation for responsible ai,” inProceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 2022

    2022

  32. [32]

    Transnet v2: An effective deep network architecture for fast shot transition detection,

    T. Souček and J. Lokoč, “Transnet v2: An effective deep network architecture for fast shot transition detection,”arXiv preprint arXiv:2008.04838, 2020

    Pith/arXiv arXiv 2008

  33. [33]

    Dense-captioning events in videos,

    R. Krishna, K. Hata, F. Ren, L. Fei-Fei, and J. C. Niebles, “Dense-captioning events in videos,” inProceedings of the IEEE 12 Visko Orbis 1.0 International Conference on Computer Vision, 2017

    2017

  34. [34]

    Vid2seq: Large-scale pretraining of a visual language model for dense video captioning,

    A. Yang, A. Nagrani, P. H. Seo, A. Miech, J. Pont-Tuset, I. Laptev, J. Sivic, and C. Schmid, “Vid2seq: Large-scale pretraining of a visual language model for dense video captioning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023

  35. [35]

    Miradata: A large-scale video dataset with long durations and structured captions,

    X. Ju, Y. Gao, Z. Zhang, Z. Yuanet al., “Miradata: A large-scale video dataset with long durations and structured captions,” arXiv preprint arXiv:2407.06358, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Flow matching for generative modeling,

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

    2023

  37. [37]

    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,” in International Conference on Learning Representations, 2023

    2023

  38. [38]

    Echo-memory: A controlled study of memory in action world models,

    W. King, Z. Xue, Y. Bian, J. Huanget al., “Echo-memory: A controlled study of memory in action world models,”arXiv preprint arXiv:2606.09803, 2026

    Pith/arXiv arXiv 2026

  39. [39]

    Diffusion forcing: Next-token prediction meets full-sequence diffusion,

    B. Chen, D. Marti Monso, Y. Du, M. Simchowitz, R. Tedrake, and V. Sitzmann, “Diffusion forcing: Next-token prediction meets full-sequence diffusion,” inAdvances in Neural Information Processing Systems, 2024

    2024

  40. [40]

    On distillation of guided diffusion models,

    C. Meng, R. Rombach, R. Gao, D. Kingma, S. Ermon, J. Ho, and T. Salimans, “On distillation of guided diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 14297–14306

    2023

  41. [41]

    Consistency models,

    Y. Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency models,” inInternational Conference on Machine Learning, 2023

    2023

  42. [42]

    Large scale diffusion distillation via score-regularized continuous-time consistency,

    K. Zheng, Y. Wang, Q. Ma, H. Chen, J. Zhang, Y. Balaji, J. Chen, M.-Y. Liu, J. Zhu, and Q. Zhang, “Large scale diffusion distillation via score-regularized continuous-time consistency,” inInternational Conference on Learning Representations,

  43. [43]

    Causal-rcm: A unified teacher-forcing and self-forcing open recipe for autoregressive diffusion distillation in streaming video generation and interactive world models,

    K. Zheng, G. He, M. Zhao, J. Zhang, H. Chen, J. Chen, C.-H. Lin, M.-Y. Liu, J. Zhu, and Q. Ma, “Causal-rcm: A unified teacher-forcing and self-forcing open recipe for autoregressive diffusion distillation in streaming video generation and interactive world models,”arXiv preprint arXiv:2606.25473, 2026

    Pith/arXiv arXiv 2026

  44. [44]

    One-step diffusion with distribution matching distillation,

    T. Yin, M. Gharbi, R. Zhang, E. Shechtman, F. Durand, W. T. Freeman, and T. Park, “One-step diffusion with distribution matching distillation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 6613–6623

    2024

  45. [45]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models,

    Z. Shao, P. Wang, Q. Zhu, R. Xuet al., “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,”arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  46. [46]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

  47. [47]

    Flow-grpo: Training flow matching models via online rl,

    J. Liu, G. Liu, J. Liang, Y. Li, J. Liu, X. Wang, P. Wan, D. Zhang, and W. Ouyang, “Flow-grpo: Training flow matching models via online rl,”arXiv preprint arXiv:2505.05470, 2025

    Pith/arXiv arXiv 2025

  48. [48]

    Dancegrpo: Unleashing grpo on visual generation,

    Z. Xue, J. Wu, Y. Gao, F. Kong, L. Zhu, M. Chen, Z. Liu, W. Liu, Q. Guo, W. Huang, and P. Luo, “Dancegrpo: Unleashing grpo on visual generation,”arXiv preprint arXiv:2505.07818, 2025

    Pith/arXiv arXiv 2025

  49. [49]

    Deep compression autoencoder for efficient high-resolution diffusion models,

    J. Chen, H. Cai, J. Chen, E. Xie, S. Yang, H. Tang, M. Li, Y. Lu, and S. Han, “Deep compression autoencoder for efficient high-resolution diffusion models,”arXiv preprint arXiv:2410.10733, 2024

    Pith/arXiv arXiv 2024

  50. [50]

    Turbo-vaed: Fast and stable transfer of video-vaes to mobile devices,

    Y. Zou, J. Yao, S. Yu, S. Zhang, W. Liu, and X. Wang, “Turbo-vaed: Fast and stable transfer of video-vaes to mobile devices,”arXiv preprint arXiv:2508.09136, 2025

    Pith/arXiv arXiv 2025

  51. [51]

    Flash-vaed: Plug-and-play vae decoders for efficient video generation,

    L. Zhu, Y. Huang, X. Ge, Y. Xueet al., “Flash-vaed: Plug-and-play vae decoders for efficient video generation,”arXiv preprint arXiv:2602.19161, 2026

    arXiv 2026

  52. [52]

    Image quality assessment: Unifying structure and texture similarity,

    K. Ding, K. Ma, S. Wang, and E. P. Simoncelli, “Image quality assessment: Unifying structure and texture similarity,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021

    2021

  53. [53]

    Frame-recurrent video super-resolution,

    M. S. M. Sajjadi, R. Vemulapalli, and M. Brown, “Frame-recurrent video super-resolution,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 6626–6634

    2018

  54. [54]

    Learning temporal coherence via self-supervision for gan-based video generation,

    M. Chu, Y. Xie, J. Mayer, L. Leal-Taixé, and N. Thuerey, “Learning temporal coherence via self-supervision for gan-based video generation,”ACM Transactions on Graphics, 2020

    2020

  55. [55]

    RAFT: Recurrent all-pairs field transforms for optical flow,

    Z. Teed and J. Deng, “RAFT: Recurrent all-pairs field transforms for optical flow,” inEuropean Conference on Computer Vision, 2020

    2020

  56. [56]

    VRT: A video restoration transformer,

    J. Liang, J. Cao, Y. Fan, K. Zhang, R. Ranjan, Y. Li, R. Timofte, and L. Van Gool, “VRT: A video restoration transformer,” arXiv preprint arXiv:2201.12288, 2022

    Pith/arXiv arXiv 2022

  57. [57]

    Recurrent video restoration transformer with guided deformable attention,

    J. Liang, Y. Fan, X. Xiang, R. Ranjan, E. Ilg, S. Green, J. Cao, K. Zhang, R. Timofte, and L. Van Gool, “Recurrent video restoration transformer with guided deformable attention,” inAdvances in Neural Information Processing Systems, 2022

    2022

  58. [58]

    BasicVSR++: Improving video super-resolution with enhanced propagation and alignment,

    K. C. K. Chan, S. Zhou, X. Xu, and C. C. Loy, “BasicVSR++: Improving video super-resolution with enhanced propagation and alignment,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 5972–5981

    2022

  59. [59]

    Roformer: Enhanced transformer with rotary position embedding,

    J. Su, Y. Lu, S. Pan, A. Murtadha, B. Wen, and Y. Liu, “Roformer: Enhanced transformer with rotary position embedding,” arXiv preprint arXiv:2104.09864, 2021

    Pith/arXiv arXiv 2021

  60. [60]

    Prompt cache: Modular attention reuse for low-latency inference,

    I. Gim, G. Chen, S.-s. Lee, N. Sarda, A. Khandelwal, and L. Zhong, “Prompt cache: Modular attention reuse for low-latency inference,”arXiv preprint arXiv:2311.04934, 2023

    Pith/arXiv arXiv 2023

  61. [61]

    SGLang: Efficient execution of structured language model programs,

    L. Zheng, L. Yin, Z. Xie, C. Sun, J. Huang, C. H. Yu, S. Cao, C. Kozyrakis, I. Stoica, J. E. Gonzalez, C. Barrett, and Y. Sheng, “SGLang: Efficient execution of structured language model programs,”arXiv preprint arXiv:2312.07104, 2024. 13 Visko Orbis 1.0

    Pith/arXiv arXiv 2024

  62. [62]

    Longlive-2.0: An nvfp4 parallel infrastructure for long video generation,

    Y. Chen, L. Wang, W. Huang, S. Yang, B. Zhang, Y. Xiao, R. Chu, W. Mao, Q. Hu, S. Liu, Y. Zhao, H. Mao, Y.-C. Chen, E. Xie, X. Qi, and S. Han, “Longlive-2.0: An nvfp4 parallel infrastructure for long video generation,”arXiv preprint arXiv:2605.18739, 2026

    Pith/arXiv arXiv 2026

  63. [63]

    TVM: An automated end-to-end optimizing compiler for deep learning,

    T. Chen, T. Moreau, Z. Jiang, L. Zheng, E. Yan, H. Shen, M. Cowan, L. Wang, Y. Hu, L. Ceze, C. Guestrin, and A. Krishnamurthy, “TVM: An automated end-to-end optimizing compiler for deep learning,” in13th USENIX Symposium on Operating Systems Design and Implementation, 2018, pp. 578–594

    2018

  64. [64]

    Triton: An intermediate language and compiler for tiled neural network computations,

    P. Tillet, H. T. Kung, and D. Cox, “Triton: An intermediate language and compiler for tiled neural network computations,” inProceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages, 2019, pp. 10–19

    2019

  65. [65]

    PyTorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation,

    J. Ansel, E. Yang, H. He, N. Gimelshein, A. Jain, M. Voznesensky, B. Bao, P. Bell, D. Berard, E. Burovskiet al., “PyTorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation,” inProceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2...

    2024

  66. [66]

    Flashattention-2: Faster attention with better parallelism and work partitioning,

    T. Dao, “Flashattention-2: Faster attention with better parallelism and work partitioning,” inInternational Conference on Learning Representations, 2024

    2024

  67. [67]

    UniPC: A unified predictor-corrector framework for fast sampling of diffusion models,

    W. Zhao, L. Bai, Y. Rao, J. Zhou, and J. Lu, “UniPC: A unified predictor-corrector framework for fast sampling of diffusion models,”arXiv preprint arXiv:2302.04867, 2023

    Pith/arXiv arXiv 2023

  68. [68]

    Smoothquant: Accurate and efficient post-training quantization for large language models,

    G. Xiao, J. Lin, M. Seznec, H. Wu, J. Demouth, and S. Han, “Smoothquant: Accurate and efficient post-training quantization for large language models,” inInternational Conference on Machine Learning, 2023

    2023

  69. [69]

    ViDiT-Q: Efficient and accurate quantization of diffusion transformers for image and video generation,

    T. Zhao, T. Fang, E. Liu, R. Wan, W. Soedarmadji, S. Li, Z. Lin, G. Dai, S. Yan, H. Yang, X. Ning, and Y. Wang, “ViDiT-Q: Efficient and accurate quantization of diffusion transformers for image and video generation,” inInternational Conference on Learning Representations, 2025

    2025

  70. [70]

    Deepspeed ulysses: System optimizations for enabling training of extreme long sequence transformer models,

    S. A. Jacobs, M. Tanaka, C. Zhang, M. Zhang, S. L. Song, S. Rajbhandari, and Y. He, “Deepspeed ulysses: System optimizations for enabling training of extreme long sequence transformer models,”arXiv preprint arXiv:2309.14509, 2023

    Pith/arXiv arXiv 2023

  71. [71]

    USP: A unified sequence parallelism approach for long context generative ai,

    J. Fang and S. Zhao, “USP: A unified sequence parallelism approach for long context generative ai,”arXiv preprint arXiv:2405.07719, 2024

    Pith/arXiv arXiv 2024

  72. [72]

    xDiT: An inference engine for diffusion transformers with massive parallelism,

    J. Fang, J. Pan, X. Sun, A. Li, and J. Wang, “xDiT: An inference engine for diffusion transformers with massive parallelism,” arXiv preprint arXiv:2411.01738, 2024

    Pith/arXiv arXiv 2024

  73. [73]

    Wf-vae: Enhancing video vae by wavelet-driven energy flow for latent video diffusion model,

    Z. Li, B. Lin, Y. Ye, L. Chen, X. Cheng, S. Yuan, and L. Yuan, “Wf-vae: Enhancing video vae by wavelet-driven energy flow for latent video diffusion model,”arXiv preprint arXiv:2411.17459, 2024

    Pith/arXiv arXiv 2024

  74. [74]

    RIFLEx: A free lunch for length extrapolation in video diffusion transformers,

    M. Zhao, G. He, Y. Chen, H. Zhu, C. Li, and J. Zhu, “RIFLEx: A free lunch for length extrapolation in video diffusion transformers,”arXiv preprint arXiv:2502.15894, 2025

    Pith/arXiv arXiv 2025

  75. [75]

    Inference-time physics alignment of video generative models with latent world models,

    J. Yuan, X. Zhang, F. Friedrich, N. Beltran-Velez, M. Hall, R. Askari-Hemmat, X. Han, N. Ballas, M. Drozdzal, and A. Romero-Soriano, “Inference-time physics alignment of video generative models with latent world models,”arXiv preprint arXiv:2601.10553, 2026

    arXiv 2026

  76. [76]

    Online normalizer calculation for softmax,

    M. Milakov and N. Gimelshein, “Online normalizer calculation for softmax,”arXiv preprint arXiv:1805.02867, 2018

    Pith/arXiv arXiv 2018

  77. [77]

    Narrlv: Towards a comprehensive narrative-centric evaluation for long video generation models,

    X. Fenget al., “Narrlv: Towards a comprehensive narrative-centric evaluation for long video generation models,”arXiv preprint arXiv:2507.11245, 2025

    arXiv 2025

  78. [78]

    Directorbench: Diagnosing long-form video generation with personalized multi-agent evaluation,

    J. Chen, Q. Chen, J. Zhang, Y. Wuet al., “Directorbench: Diagnosing long-form video generation with personalized multi-agent evaluation,”arXiv preprint arXiv:2605.30090, 2026

    Pith/arXiv arXiv 2026

  79. [79]

    Infinitystar: Unified spacetime autoregressive modeling for visual generation,

    J. Liu, J. Han, B. Yan, H. Wu, F. Zhu, X. Wang, Y. Jiang, B. Peng, and Z. Yuan, “Infinitystar: Unified spacetime autoregressive modeling for visual generation,”arXiv preprint arXiv:2511.04675, 2025

    arXiv 2025

  80. [80]

    Sana-video: Efficient video generation with block linear diffusion transformer,

    J. Chen, Y. Zhao, J. Yuet al., “Sana-video: Efficient video generation with block linear diffusion transformer,”arXiv preprint arXiv:2509.24695, 2025

    arXiv 2025

Showing first 80 references.