pith. sign in

arxiv: 2605.16649 · v1 · pith:75VFFQ55new · submitted 2026-05-15 · 💻 cs.CV

AtlasVid: Efficient Ultra-High-Resolution Long Video Generation via Decoupled Global-Local Modeling

Ziyang Mai , Yuyao Zhang , Yu-Wing Tai This is my paper

Pith reviewed 2026-05-20 18:20 UTC · model grok-4.3

classification 💻 cs.CV
keywords video generationdiffusion modelsultra-high-resolutionlong video synthesisdecoupled modelingglobal-local attentionRoPE scaling
0
0 comments X

The pith

Decoupled global-local modeling trains video generators at low resolution to produce ultra-high-resolution long videos.

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

The paper argues that video diffusion models already hold strong local visual priors, so the real barrier to ultra-high-resolution long videos is extending global spatiotemporal coherence without exploding compute costs. AtlasVid addresses this by first creating a low-resolution low-FPS global semantic proxy using temporally scaled RoPE, then guiding a high-resolution detail branch through joint denoising with hierarchical locality-preserving attention. This separation lets the model train only at 720P with lightweight LoRA adaptation yet generalize directly to 4K and beyond for sequences longer than 10 seconds. The design claims both a 60.9x speedup and higher quality than generators trained natively at full resolution.

Core claim

Existing video diffusion models encode strong local priors; the bottleneck is efficient global modeling at scale. AtlasVid therefore decouples the problem: a global branch generates a low-resolution low-FPS semantic proxy via temporally scaled RoPE to extend temporal horizon without raising token count, while a high-resolution branch performs joint denoising under reordered spatiotemporal windows and asymmetric global-local attention that injects aligned guidance while preserving pretrained local ability.

What carries the argument

Temporally scaled RoPE global semantic proxy that guides joint denoising in a high-resolution branch equipped with hierarchical locality-preserving attention.

If this is right

  • Training occurs only at 720P yet the model directly synthesizes 4K videos longer than 10 seconds without full retraining.
  • Generation runs 60.9 times faster than native high-resolution approaches while using less training compute.
  • Quality exceeds that of models trained from scratch at 4K because local priors remain untouched.
  • The framework supports resolution-agnostic deployment for arbitrary output sizes after a single low-resolution training run.

Where Pith is reading between the lines

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

  • The same proxy-plus-detail split could be tested on other generative tasks where global structure and local texture must be handled at different scales.
  • If the proxy guidance proves robust, the method could lower the barrier to creating custom high-resolution video models without access to large native 4K datasets.
  • Extending the temporal scaling factor in RoPE might allow even longer coherent videos without further increases in memory footprint.

Load-bearing premise

The low-resolution low-FPS global proxy supplies enough aligned semantic guidance that joint denoising can keep both long-range temporal coherence and fine spatial details intact.

What would settle it

Generate the same prompt at 4K with the proposed method and with a native 4K baseline; if the decoupled outputs show visibly broken motion continuity or missing fine detail while the native baseline does not, the sufficiency of the low-res proxy is refuted.

Figures

Figures reproduced from arXiv: 2605.16649 by Yu-Wing Tai, Yuyao Zhang, Ziyang Mai.

Figure 1
Figure 1. Figure 1: AtlasVid enables the generation of ultra-high-resolution and long-duration videos in different settings, including 8K 29 frames, 4K 161 frames and 2K 321 frames. Frame index indicated in the top-left corner and the output resolution in the top-right corner. Bottom: AtlasVid runs 60.9× faster than UltraWan at 4K × 81 frames (left) and reduces per-layer attention FLOPs by up to 1208.2× over FlashAttention fr… view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline of AtlasVid . It first employs a semantic generator to produce a low-resolution, low-frame-rate video that serves as a global semantic proxy. Conditioned on this reference, the second stage performs spatiotemporal detail generation through an efficient hierarchical locality￾preserving attention mechanism, enabling ultra-high-resolution long-video synthesis(UHRL video) with substantially improved c… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of long ultra-high-resolution video generation. Top: The first two [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation on the importance of our attention design. The first two columns demonstrate the [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation study on 4K data finetuning. With 4K finetuning (top), the model produces more realistic fine￾grained details, while without 4K finetuning (bottom) it can still generate plausible details, demonstrating the robustness of our base model. Ablation on 4K data fine-tuning [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: 4K 161 Frames results: Each results spans one row. Examples show no quality degradation [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: 4K 161 Frames results: Each results spans one row. Examples show no quality degradation [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: 2K 321 Frames results: Each results spans for two rows. The examples here shows large [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: 2K 321 Frames results: Each results spans for two rows. The examples here shows large [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: 8K 29 Frames results: Each results spans for two rows. The frame indices are 0 and 28. [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
read the original abstract

Recent diffusion-based video generators have achieved remarkable visual fidelity and prompt controllability, yet scaling them to ultra-high-resolution (UHR) long videos remains prohibitively expensive. The difficulty is especially pronounced for long single-shot generation where a continuous scene must preserve global temporal coherence, and fine-grained spatial details without relying on clip transitions or autoregressive shot stitching. In this work, we revisit this challenge from the perspective of decoupled modeling. We argue that existing video diffusion models already encode strong local visual priors, while the main bottleneck lies in efficiently extending global spatiotemporal modeling as resolution and duration increase. Based on this insight, we propose AtlaVid, a decoupled global-local framework for efficient UHR long video generation. AtlaVid first generates a low-resolution and low-FPS global semantic proxy via temporally scaled RoPE, thereby extending the temporal horizon without increasing the training token count. Guided by this proxy, a high-resolution detail branch performs joint denoising with hierarchical locality-preserving attention. Reordered spatiotemporal windows preserve geometric locality and asymmetric global-local attention injects aligned semantic guidance and preserves the model's pretrained ability. This design enables resolution-agnostic training: the model is trained only at 720P with lightweight LoRA adaptation, yet generalizes directly to 4K and beyond for longer (>10s) video synthesis. Experiments show that AtlaVid substantially improves the efficiency of ultra-high-resolution long video generation, achieving high-quality UHR long video generation with 60.9x speed up and significantly less training cost and even better performance than native 4K video generators.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes AtlasVid, a decoupled global-local framework for efficient ultra-high-resolution long video generation. It generates a low-resolution low-FPS global semantic proxy using temporally scaled RoPE to extend the temporal horizon without increasing token count, then performs joint denoising in a high-resolution detail branch guided by hierarchical locality-preserving attention and asymmetric global-local attention. The design claims to enable resolution-agnostic training at 720P with lightweight LoRA adaptation that generalizes directly to 4K and beyond for videos longer than 10s, achieving a 60.9x speedup and better performance than native 4K generators while preserving global temporal coherence and fine spatial details.

Significance. If the empirical claims hold, the work would be significant for video diffusion models by demonstrating a practical path to scale beyond current resolution and duration limits without full high-resolution retraining. The emphasis on reusing pretrained local priors and decoupling global semantics via a proxy could reduce compute barriers in the field, provided the guidance mechanism proves robust.

major comments (3)
  1. [Abstract] Abstract: the central generalization claim (resolution-agnostic training at 720P generalizing to 4K with 60.9x speedup) rests on unshown experiments; no quantitative metrics, baselines, error bars, or ablation details are supplied to support the speedup or coherence preservation over >10s sequences.
  2. [Method] Method description: the temporally scaled RoPE proxy operates at reduced FPS while the detail branch performs joint denoising at native 4K; no equations or analysis demonstrate that upsampled guidance from the low-FPS proxy maintains frame-to-frame temporal coherence without drift or hallucination in long continuous shots.
  3. [Experiments] Experiments section: the claim that reordered spatiotemporal windows and hierarchical attention preserve both global coherence and fine details without high-resolution training data requires explicit ablations on proxy FPS/resolution and quantitative comparisons against native 4K baselines to be load-bearing for the resolution-agnostic assertion.
minor comments (2)
  1. [Method] Clarify notation for 'temporally scaled RoPE' versus standard RoPE in the method section to avoid ambiguity in how temporal scaling is implemented.
  2. [Discussion] Add a short discussion of potential failure modes when the low-FPS proxy provides insufficient granularity for very long shots.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point-by-point below. We agree that additional clarity, quantitative details, and analysis will strengthen the manuscript and have revised accordingly where the comments identify gaps in presentation or supporting evidence.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central generalization claim (resolution-agnostic training at 720P generalizing to 4K with 60.9x speedup) rests on unshown experiments; no quantitative metrics, baselines, error bars, or ablation details are supplied to support the speedup or coherence preservation over >10s sequences.

    Authors: We acknowledge that the abstract, being a concise summary, does not contain the full experimental details. The quantitative metrics supporting the 60.9x speedup (measured as the ratio of wall-clock inference time on identical hardware for equivalent-length 4K outputs), coherence metrics over sequences longer than 10s, baseline comparisons, and error bars from repeated runs are reported in Section 4 (Experiments) and the associated tables/figures. To address the concern directly, we will revise the abstract to include a brief parenthetical reference to these results and the specific sections where they appear, ensuring the central claims are transparently linked to the supporting evidence without exceeding abstract length constraints. revision: partial

  2. Referee: [Method] Method description: the temporally scaled RoPE proxy operates at reduced FPS while the detail branch performs joint denoising at native 4K; no equations or analysis demonstrate that upsampled guidance from the low-FPS proxy maintains frame-to-frame temporal coherence without drift or hallucination in long continuous shots.

    Authors: We agree that the current method description would benefit from explicit equations and analysis on temporal coherence. In the revised manuscript we will add a new subsection (or expanded paragraph) in Section 3 that includes: (1) the mathematical formulation of temporally scaled RoPE and the upsampling operator from the low-FPS proxy to the high-resolution branch; (2) a short derivation showing how the asymmetric global-local attention and locality-preserving mechanism align semantic guidance across frames; and (3) an empirical coherence analysis (e.g., frame-to-frame optical flow consistency and drift metrics) on long continuous shots. These additions will directly demonstrate the absence of drift or hallucination under the proposed guidance. revision: yes

  3. Referee: [Experiments] Experiments section: the claim that reordered spatiotemporal windows and hierarchical attention preserve both global coherence and fine details without high-resolution training data requires explicit ablations on proxy FPS/resolution and quantitative comparisons against native 4K baselines to be load-bearing for the resolution-agnostic assertion.

    Authors: We appreciate this observation. While the current experiments section contains comparisons to native 4K generators and some attention-related ablations, we concur that more targeted ablations on proxy FPS and resolution are needed to make the resolution-agnostic claim fully load-bearing. In the revision we will add a dedicated ablation study (new table or figure) that systematically varies proxy FPS (e.g., 1, 2, 4 fps) and resolution, reporting quantitative metrics including FVD, temporal coherence scores, and direct side-by-side comparisons against native 4K training baselines. This will provide the explicit evidence requested. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's core derivation introduces independent architectural elements including temporally scaled RoPE for low-resolution low-FPS global proxy generation and hierarchical locality-preserving attention with reordered spatiotemporal windows for the high-resolution detail branch. These choices are presented as design decisions that enable resolution-agnostic training at 720P with LoRA adaptation and direct generalization to 4K, without any equations or steps that reduce the claimed speedup, coherence preservation, or performance metrics back to fitted parameters or quantities extracted from the target high-resolution outputs by construction. No self-citation chains, uniqueness theorems, or ansatzes are invoked in a load-bearing way that collapses the argument to prior author work or tautological renaming; the decoupling insight and proxy-guidance mechanism stand as self-contained modeling assumptions whose validity is left to empirical validation rather than definitional equivalence.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The approach rests on standard diffusion model assumptions and the untested premise that a low-resolution proxy suffices for guidance; no new physical entities or ad-hoc constants are introduced in the abstract.

axioms (2)
  • domain assumption Existing video diffusion models already encode strong local visual priors
    Stated directly in the abstract as the basis for focusing compute on global modeling.
  • domain assumption Temporally scaled RoPE extends temporal horizon without increasing token count
    Core technical assumption enabling the global proxy.

pith-pipeline@v0.9.0 · 5820 in / 1363 out tokens · 78502 ms · 2026-05-20T18:20:45.261972+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages · 12 internal anchors

  1. [1]

    Kandinsky 5.0: A family of foundation models for image and video generation.arXiv preprint arXiv:2511.14993,

    Vladimir Arkhipkin, Vladimir Korviakov, Nikolai Gerasimenko, Denis Parkhomenko, Viacheslav Vasilev, Alexey Letunovskiy, Nikolai Vaulin, Maria Kovaleva, Ivan Kirillov, Lev Novitskiy, et al. Kandinsky 5.0: A family of foundation models for image and video generation.arXiv preprint arXiv:2511.14993,

    work page arXiv

  2. [2]

    SkyReels-V2: Infinite-length Film Generative Model

    Guibin Chen, Dixuan Lin, Jiangping Yang, Chunze Lin, Junchen Zhu, Mingyuan Fan, Hao Zhang, Sheng Chen, Zheng Chen, Chengcheng Ma, et al. Skyreels-v2: Infinite-length film generative model.arXiv preprint arXiv:2504.13074, 2025a. Junsong Chen, Yuyang Zhao, Jincheng Yu, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Daquan Zhou, Huan Ling, ...

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning

    Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning.arXiv preprint arXiv:2307.08691,

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    google/technologies/veo/

    URL https://deepmind. google/technologies/veo/. Zhentao Fan, Zongzuo Wang, and Weiwei Zhang. Taocache: Structure-maintained video generation acceleration.arXiv preprint arXiv:2508.08978,

    work page arXiv

  5. [5]

    LTX-Video: Realtime Video Latent Diffusion

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    LTX-2: Efficient Joint Audio-Visual Foundation Model

    Yoav HaCohen, Benny Brazowski, Nisan Chiprut, Yaki Bitterman, Andrew Kvochko, Avishai Berkowitz, Daniel Shalem, Daphna Lifschitz, Dudu Moshe, Eitan Porat, et al. Ltx-2: Efficient joint audio-visual foundation model.arXiv preprint arXiv:2601.03233,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    Venhancer: Generative space-time enhancement for video generation

    Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. Venhancer: Generative space-time enhancement for video generation.arXiv preprint arXiv:2407.07667,

    work page arXiv

  8. [8]

    CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers

    Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers.arXiv preprint arXiv:2205.15868,

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    In-context lora for diffusion transformers.arXiv preprint arXiv:2410.23775, 2024a

    Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers.arXiv preprint arXiv:2410.23775, 2024a. Xun Huang, Zhengqi Li, Guande He, Mingyuan Zhou, and Eli Shechtman. Self forcing: Bridging the train-test gap in autoregressive video diffusion.arXiv preprint...

    work page arXiv

  10. [10]

    HunyuanVideo: A Systematic Framework For Large Video Generative Models

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603,

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Skyreels-v3 technique report.arXiv preprint arXiv:2601.17323,

    Debang Li, Zhengcong Fei, Tuanhui Li, Yikun Dou, Zheng Chen, Jiangping Yang, Mingyuan Fan, Jingtao Xu, Jiahua Wang, Baoxuan Gu, et al. Skyreels-v3 technique report.arXiv preprint arXiv:2601.17323,

    work page arXiv

  12. [12]

    Open-Sora Plan: Open-Source Large Video Generation Model

    Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Timestep embedding tells: It’s time to cache for video diffusion model, 2024

    Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. Timestep embedding tells: It’s time to cache for video diffusion model. arXiv preprint arXiv:2411.19108,

    work page arXiv

  14. [14]

    Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference.arXiv preprint arXiv:2310.04378,

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    Step-Video-T2V Technical Report: The Practice, Challenges, and Future of Video Foundation Model

    Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiaoniu Song, Xing Chen, et al. Step-video-t2v technical report: The practice, challenges, and future of video foundation model.arXiv preprint arXiv:2502.10248, 2025a. Zehong Ma, Longhui Wei, Feng Wang, Shiliang Zhang, and Qi Tian. Magcache: Fast video ge...

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Cinescale: Free lunch in high-resolution cinematic visual generation.arXiv preprint arXiv:2508.15774, 2025

    Haonan Qiu, Ning Yu, Ziqi Huang, Paul Debevec, and Ziwei Liu. Cinescale: Free lunch in high- resolution cinematic visual generation.arXiv preprint arXiv:2508.15774,

    work page arXiv

  17. [17]

    HunyuanVideo 1.5 Technical Report

    URL https: //arxiv.org/abs/2511.18870. Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. InEuropean conference on computer vision, pages 402–419. Springer,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    MAGI-1: Autoregressive Video Generation at Scale

    Hansi Teng, Hongyu Jia, Lei Sun, Lingzhi Li, Maolin Li, Mingqiu Tang, Shuai Han, Tianning Zhang, WQ Zhang, Weifeng Luo, et al. Magi-1: Autoregressive video generation at scale.arXiv preprint arXiv:2505.13211,

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    Wan: Open and Advanced Large-Scale Video Generative Models

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314,

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    Freeswim: Revisiting sliding- window attention mechanisms for training-free ultra-high-resolution video generation.arXiv preprint arXiv:2511.14712,

    Yunfeng Wu, Jiayi Song, Zhenxiong Tan, Zihao He, and Songhua Liu. Freeswim: Revisiting sliding- window attention mechanisms for training-free ultra-high-resolution video generation.arXiv preprint arXiv:2511.14712,

    work page arXiv

  21. [21]

    Ultravideo: High-quality uhd video dataset with comprehensive captions.arXiv preprint arXiv:2506.13691, 2025

    Zhucun Xue, Jiangning Zhang, Teng Hu, Haoyang He, Yinan Chen, Yuxuan Cai, Yabiao Wang, Chengjie Wang, Yong Liu, Xiangtai Li, et al. Ultravideo: High-quality uhd video dataset with comprehensive captions.arXiv preprint arXiv:2506.13691,

    work page arXiv

  22. [22]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613–6623, 2024a. Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli ...

    work page arXiv

  23. [23]

    Luve: Latent-cascaded ultra-high-resolution video generation with dual frequency experts.arXiv preprint arXiv:2602.11564,

    Chen Zhao, Jiawei Chen, Hongyu Li, Zhuoliang Kang, Shilin Lu, Xiaoming Wei, Kai Zhang, Jian Yang, and Ying Tai. Luve: Latent-cascaded ultra-high-resolution video generation with dual frequency experts.arXiv preprint arXiv:2602.11564,

    work page arXiv

  24. [24]

    Videogen-of-thought: Step-by-step generating multi-shot video with minimal manual intervention, 2025

    Mingzhe Zheng, Yongqi Xu, Haojian Huang, Xuran Ma, Yexin Liu, Wenjie Shu, Yatian Pang, Feilong Tang, Qifeng Chen, Harry Yang, et al. Videogen-of-thought: A collaborative framework for multi-shot video generation.arXiv preprint arXiv:2412.02259, 3(6),

    work page arXiv

  25. [25]

    Flashvsr: Towards real- time diffusion-based streaming video super-resolution.arXiv preprint arXiv:2510.12747, 2025

    Junhao Zhuang, Shi Guo, Xin Cai, Xiaohui Li, Yihao Liu, Chun Yuan, and Tianfan Xue. Flashvsr: To- wards real-time diffusion-based streaming video super-resolution.arXiv preprint arXiv:2510.12747,

    work page arXiv

  26. [26]

    Mixed-precision (bf16) is used throughout, and we adopt a flow-matching objective consistent with the Wan2.1 base model

    for 15K iterations. Mixed-precision (bf16) is used throughout, and we adopt a flow-matching objective consistent with the Wan2.1 base model. Stage 1 (semantic generator).We finetune the base model with temporal-scale RoPE ( rt=4) on 720P×81 -frame clips sub-sampled at 4 fps, so that an 81-frame proxy spans an effective horizon of ∼20seconds at16fps target...

    work page 2025

  27. [27]

    C Detailed Metric Definitions This section formalises the metrics referenced in Table 2 and the ablation tables of the main paper

    End-to-end runtime includes text encoding, both denoising stages, and 3D-V AE decoding. C Detailed Metric Definitions This section formalises the metrics referenced in Table 2 and the ablation tables of the main paper. For all metrics, v∈R T×H×W×3 denotes a video. Higher-is-better metrics are marked ↑ and lower-is-better↓. C.1 High-Definition Metrics The ...

    work page 2026