pith. sign in

arxiv: 2606.09150 · v1 · pith:7YQP7MVAnew · submitted 2026-06-08 · 💻 cs.CV

Ultra Flash: Scaling Real-Time Streaming Video Generation to High Resolutions

Luxury , Jie Huang , Zihao Fan , Xiaoxiao Ma , Yuming Li , Jun-hao Zhuang , Zeyue Xue , Siming Fu
show 20 more authors
Haoran Li Mingchen Zhong Guohui Zhang Shichen Ma Yijun Liu Jiaqi Shi Yanwen Ma Yaofeng Su Haoyu Wang Yaowei Li Songchun Zhang Weiyang Jin Yuxuan Bian Shiyi Zhang Haojun Xu Shuai Lu Xin Han Wei Tang Haoyang Huang Nan Duan
This is my paper

Pith reviewed 2026-06-27 17:06 UTC · model grok-4.3

classification 💻 cs.CV
keywords real-time video generationstreaming videosuper-resolutionvideo diffusion modelscascaded frameworkhigh-resolution videolatent upsamplerpreference optimization
0
0 comments X

The pith

Ultra Flash scales low-resolution streaming video models to real-time 1K and 2K output on a single GPU via a cascaded super-resolution framework.

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

The paper introduces Ultra Flash, a cascaded framework that extends existing low-resolution autoregressive video diffusion models to high resolutions while preserving real-time speeds. It relies on an architecture-preserving training method for the super-resolution stage, a causal latent upsampler, and a multi-stage optimization process that includes distillation and preference tuning. A sympathetic reader would care because prior models were limited to low resolutions such as 480P, restricting practical real-time use. If the claims hold, high-resolution streaming generation becomes feasible on consumer hardware without major quality loss.

Core claim

Ultra Flash is a cascaded streaming framework for real-time high-resolution video generation. It achieves approximately 30 FPS at 1K resolution and 18 FPS at 2K resolution on a single GPU by training a super-resolution model in an architecture-preserving T2V-to-TV2V manner with an AIGC-oriented data degradation pipeline, pairing it with a causal streaming latent upsampler and high-resolution decoder, and applying a cascade optimization scheme of hybrid-reward sparse causalization, single-step distillation, and cascaded streaming self-forcing preference optimization with dynamic cache management, while maintaining state-of-the-art visual quality.

What carries the argument

The cascaded streaming framework that combines an architecture-preserving T2V-to-TV2V super-resolution training paradigm, a causal streaming latent upsampler with high-resolution decoder, and a cascade optimization scheme using sparse causalization, distillation, and self-forcing preference optimization.

If this is right

  • Existing low-resolution generative models can be extended to high-resolution real-time streaming output without retraining the base model from scratch.
  • Spatiotemporal coherence improves through the causal latent upsampler with only negligible added computation.
  • The optimization scheme jointly raises coherence and quality while supporting real-time performance via dynamic cache management.
  • The full pipeline produces ultra-high-resolution streaming video at the reported frame rates on one GPU.
  • The approach enables efficient latent spatial scaling followed by precise high-resolution decoding.

Where Pith is reading between the lines

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

  • The same training and cascade structure could be tested on non-video generative tasks such as image or audio synthesis to check transferability.
  • Real-time applications like live content creation or interactive simulation might adopt the method once base models are widely available.
  • Adding further cascade stages could be examined to reach 4K resolutions while monitoring speed and quality trade-offs.
  • The preservation of base-model capabilities can be directly measured by comparing prompt fidelity before and after the super-resolution stage on identical inputs.

Load-bearing premise

The architecture-preserving T2V-to-TV2V super-resolution training paradigm together with the AIGC-oriented data degradation pipeline will preserve the generative capability of the base low-resolution model while still enabling enhanced high-resolution detail when cascaded.

What would settle it

Integrate the trained super-resolution stage after a standard low-resolution streaming video model and check whether the high-resolution outputs retain the base model's text prompt adherence and motion coherence or instead introduce visible artifacts and quality degradation.

Figures

Figures reproduced from arXiv: 2606.09150 by Guohui Zhang, Haojun Xu, Haoran Li, Haoyang Huang, Haoyu Wang, Jiaqi Shi, Jie Huang, Jun-hao Zhuang, Luxury, Mingchen Zhong, Nan Duan, Shichen Ma, Shiyi Zhang, Shuai Lu, Siming Fu, Songchun Zhang, Wei Tang, Weiyang Jin, Xiaoxiao Ma, Xin Han, Yanwen Ma, Yaofeng Su, Yaowei Li, Yijun Liu, Yuming Li, Yuxuan Bian, Zeyue Xue, Zihao Fan.

Figure 1
Figure 1. Figure 1: (a) Ultra Flash framework. (b) Quality–speed comparison with prior methods. Ultra Flash scales to 1K and 2K resolution while achieving better quality and real-time throughput self-forcing preference optimization with dynamic cache management: the low-resolution generator and the high-resolution SR model are jointly rolled out in a cascaded streaming fashion, where a preference optimization objective explic… view at source ↗
Figure 2
Figure 2. Figure 2: Detailed components and training of our Ultra Flash framework. (Zoom in for details.) consisting of Nb causal memory blocks, a spatial upsampling layer, a temporal expansion layer, and a channel transition convolution. The core building block is the CausalMemBlock, which fuses the current frame’s feature with the memory from the previous frame: h (ℓ) t = σ  Conv(3) 3×3  Conv(2) 3×3  σ  Conv(1) 3×3  [h… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison. Ultra Flash real-time generates sharper, temporally coherent frames at 1K & 2K while prior methods are limited to 480×832. Zoom in for the details [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison. Ultra Flash real-time generates sharper, temporally coherent frames at 1K & 2K while prior methods are limited to 480×832. Zoom in for the details [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison. Ultra Flash real-time generates sharper, temporally coherent frames at 1K & 2K while prior methods are limited to 480×832. Zoom in for the details. cache management: The three-pronged inference optimization (step reduction, IQA-adaptive cache refresh, SR cache length adaptation) significantly improves FPS with negligible quality impact. 3.4 QUALITATIVE RESULTS [PITH_FULL_IMAGE:figu… view at source ↗
Figure 6
Figure 6. Figure 6: Fine-grained texture comparison. Ultra Flash resolves high-frequency details—individual hair strands, skin texture, fabric patterns—that are lost in 480P baselines. Zoomed crops (bottom) highlight the substantial resolution advantage of our pipeline. Temporal Consistency and Color Stability [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Temporal consistency and color stability. On scenes with intricate surface details (e.g., animal skin, vegetation), Ultra Flash maintains consistent exposure, stable color, and temporally coherent textures, while baselines exhibit color drift and flickering artifacts. Complex Scene Composition [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Complex scene composition. Ultra Flash accurately renders multi-object scenes with rich spatial structure, maintaining sharp foreground details and coherent backgrounds at high resolution. Baselines produce muddled textures and lose fine structural detail. perceptual quality is preserved under dynamic conditions, while the dynamic cache management strategy maintains generation efficiency without sacrificin… view at source ↗
Figure 9
Figure 9. Figure 9: Dynamic motion and semantic coherence. Under fast camera movements and complex object interactions, Ultra Flash produces temporally smooth, artifact-free high-resolution frames, while baselines exhibit motion blur, ghosting, and temporal inconsistencies. in temporal flickering. This demonstrates that our cascaded self-forcing preference optimization effectively maintains temporal coherence during high-reso… view at source ↗
Figure 10
Figure 10. Figure 10: VBench 16-dimension radar chart. We compare Ultra Flash with Wan2.1 (teacher), CausVid, Self Forcing, and DummyForcing across all 16 VBench metrics. Ultra Flash achieves the best or near-best performance across most dimensions while maintaining real-time throughput. that our cascaded pipeline preserves dynamic motion well despite the additional SR processing, and the preference optimization prevents the m… view at source ↗
read the original abstract

While recent autoregressive video diffusion models achieve remarkable streaming quality, they remain confined to low resolutions (e.g., 480P), leaving efficient, scalable, real-time high-resolution video generation a fundamental open challenge. To bridge this gap, we present Ultra Flash, a cascaded streaming framework capable of real-time high-resolution video generation. Ultra Flash achieves ~30 FPS at 1K resolution and ~18 FPS at 2K resolution on a single GPU through three key contributions: (1) an architecture-preserving T2V-to-TV2V super-resolution training paradigm coupled with an AIGC-oriented data degradation pipeline that effectively preserves the generative capability of the base model, enabling enhanced high-resolution detail when cascaded after mainstream low-resolution generative models; (2) a causal streaming latent upsampler paired with a high-resolution decoder, which enhances spatiotemporal coherence while enabling efficient latent spatial scaling and precise high-resolution decoding with negligible computational overhead; and (3) a cascade high-resolution streaming video generation optimization scheme that first performs hybrid-reward-enhanced sparse causalization and single-step distillation of the super-resolution model, then introduces cascaded streaming self-forcing preference optimization with dynamic cache management, jointly enhancing overall coherence, improving quality, and enabling real-time high-resolution streaming video generation. Extensive experiments demonstrate that Ultra Flash reliably produces ultra-high-resolution streaming video while maintaining state-of-the-art visual quality and superior efficiency.

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

2 major / 1 minor

Summary. The paper presents Ultra Flash, a cascaded streaming framework for real-time high-resolution video generation from low-resolution T2V models. It claims ~30 FPS at 1K and ~18 FPS at 2K resolution on a single GPU with SOTA visual quality via three contributions: (1) an architecture-preserving T2V-to-TV2V super-resolution training paradigm with AIGC-oriented data degradation to preserve base model generative capability; (2) a causal streaming latent upsampler and high-resolution decoder for spatiotemporal coherence and efficient scaling; and (3) a cascade optimization scheme using hybrid-reward sparse causalization, single-step distillation, and cascaded streaming self-forcing preference optimization with dynamic cache management. Extensive experiments are said to support the efficiency and quality claims.

Significance. If the empirical results hold, this would be a meaningful advance in scaling autoregressive video diffusion models beyond low resolutions (e.g., 480P) to real-time high-res streaming, addressing a practical bottleneck. The architecture-preserving transfer and optimization techniques could be reusable in other cascaded systems. The reported single-GPU FPS figures, if substantiated with baselines and ablations, would represent a notable efficiency improvement. However, significance depends on verification of the key preservation assumption and quantitative support for the performance claims.

major comments (2)
  1. [Abstract / Contribution (1)] Abstract / Contribution (1): The claim that the T2V-to-TV2V paradigm 'effectively preserves the generative capability of the base model' is load-bearing for the cascaded real-time claim and the assertion of maintained SOTA quality. No side-by-side quantitative metrics (FVD, CLIP similarity, or streaming coherence scores) are supplied comparing the unmodified base low-resolution model to the cascaded Ultra Flash system on identical prompts. This omission leaves open whether the super-resolution stage maintains the original distribution or introduces degradation that later stages cannot fully offset.
  2. [Abstract / Contribution (3)] Abstract / Contribution (3): The optimization scheme relies on free parameters including hybrid-reward weights and distillation step count. The abstract reports no ablation studies, sensitivity analysis, or details on how these parameters were selected to achieve the stated ~30 FPS at 1K / ~18 FPS at 2K while preserving quality. This information is necessary to evaluate reproducibility and whether the performance numbers are robust.
minor comments (1)
  1. [Abstract] The abstract states performance claims and 'extensive experiments' but supplies no dataset details, baseline comparisons, error bars, or model size/hardware specifications beyond 'single GPU'. Adding these would improve clarity without altering the central claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the preservation of generative capability and the optimization details. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract / Contribution (1)] Abstract / Contribution (1): The claim that the T2V-to-TV2V paradigm 'effectively preserves the generative capability of the base model' is load-bearing for the cascaded real-time claim and the assertion of maintained SOTA quality. No side-by-side quantitative metrics (FVD, CLIP similarity, or streaming coherence scores) are supplied comparing the unmodified base low-resolution model to the cascaded Ultra Flash system on identical prompts. This omission leaves open whether the super-resolution stage maintains the original distribution or introduces degradation that later stages cannot fully offset.

    Authors: We agree that explicit side-by-side quantitative comparisons would strengthen the preservation claim. The architecture-preserving training and AIGC-oriented degradation are designed to maintain the base distribution, as supported by the overall SOTA results in our experiments. In the revision we will add a dedicated comparison table reporting FVD and CLIP similarity on identical prompts for the unmodified base model versus the cascaded system. revision: yes

  2. Referee: [Abstract / Contribution (3)] Abstract / Contribution (3): The optimization scheme relies on free parameters including hybrid-reward weights and distillation step count. The abstract reports no ablation studies, sensitivity analysis, or details on how these parameters were selected to achieve the stated ~30 FPS at 1K / ~18 FPS at 2K while preserving quality. This information is necessary to evaluate reproducibility and whether the performance numbers are robust.

    Authors: The abstract is space-constrained and summarizes contributions at a high level. Full ablation studies on hybrid-reward weights, distillation step count, and sensitivity analysis appear in Section 4.3 and the supplementary material, where we also describe the selection process that yields the reported FPS while preserving quality. We will revise the abstract to include a one-sentence reference to these ablations for improved reproducibility. revision: partial

Circularity Check

0 steps flagged

No circularity; empirical architecture and optimization claims are self-contained

full rationale

The paper advances three engineering contributions—an architecture-preserving T2V-to-TV2V training paradigm with AIGC degradation, a causal latent upsampler, and a cascade optimization scheme—validated solely through experimental results on FPS, visual quality, and coherence. No equations, fitted parameters, or predictions are presented that reduce by construction to the inputs; the central claim that the paradigm preserves base-model generative capability is an empirical assertion, not a definitional or self-citation tautology. The manuscript contains no derivation chain, uniqueness theorems, or ansatzes smuggled via prior self-work that would trigger any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central performance claims rest on the unverified effectiveness of the proposed training pipeline and optimization stages; the abstract does not enumerate hyperparameters but the method necessarily depends on many training choices (learning rates, reward weights, distillation steps, cache sizes) that function as free parameters. No new physical entities are introduced. Standard diffusion model assumptions are used without explicit listing.

free parameters (2)
  • hybrid-reward weights
    Weights balancing multiple reward signals in the sparse causalization and preference optimization stages; these are chosen to achieve the reported coherence and quality.
  • distillation step count
    Number of steps used in single-step distillation of the super-resolution model; directly affects the claimed real-time speed.
axioms (1)
  • domain assumption The base low-resolution T2V model already produces coherent streaming output that can be upscaled without retraining the entire generator.
    Invoked when the cascade is placed after mainstream low-resolution generative models.

pith-pipeline@v0.9.1-grok · 5879 in / 1432 out tokens · 23330 ms · 2026-06-27T17:06:02.028009+00:00 · methodology

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

61 extracted references · 36 canonical work pages · 18 internal anchors

  1. [1]

    Wang, Xintao and Xie, Liangbin and Dong, Chao and Shan, Ying , booktitle=. Real-

  2. [2]

    Wan: Open and Advanced Large-Scale Video Generative Models

    Wan: Open and Advanced Large-Scale Video Generative Models , author=. arXiv preprint arXiv:2503.20314 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    OmniNFT: Modality-wise Omni Diffusion Reinforcement for Joint Audio-Video Generation

    OmniNFT: Modality-wise Omni Diffusion Reinforcement for Joint Audio-Video Generation , author=. arXiv preprint arXiv:2605.12480 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer

    CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer , author=. arXiv preprint arXiv:2408.06072 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    HunyuanVideo: A Systematic Framework For Large Video Generative Models

    HunyuanVideo: A Systematic Framework For Large Video Generative Models , author=. arXiv preprint arXiv:2412.03603 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Movie Gen: A Cast of Media Foundation Models

    Movie Gen: A Cast of Media Foundation Models , author=. arXiv preprint arXiv:2410.13720 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    2024 , note=

    Video Generation Models as World Simulators , author=. 2024 , note=

    2024

  8. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    From Slow Bidirectional to Fast Autoregressive Video Diffusion Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  9. [9]

    Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion

    Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion , author=. arXiv preprint arXiv:2506.08009 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation

    Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation , author=. arXiv preprint arXiv:2602.02214 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2601.20499 , year=

    Efficient Autoregressive Video Diffusion with Dummy Head , author=. arXiv preprint arXiv:2601.20499 , year=

    work page arXiv

  12. [12]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    One-step Diffusion with Distribution Matching Distillation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  13. [13]

    arXiv preprint arXiv:2405.14867 , year=

    Improved Distribution Matching Distillation for Fast Image Synthesis , author=. arXiv preprint arXiv:2405.14867 , year=

    work page arXiv

  14. [14]

    arXiv preprint arXiv:2511.22677 , year=

    Decoupled DMD: CFG Augmentation as the Spear, Distribution Matching as the Shield , author=. arXiv preprint arXiv:2511.22677 , year=

    work page arXiv

  15. [15]

    International Conference on Machine Learning (ICML) , year=

    Consistency Models , author=. International Conference on Machine Learning (ICML) , year=

  16. [16]

    International Conference on Learning Representations (ICLR) , year=

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author=. International Conference on Learning Representations (ICLR) , year=

  17. [17]

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

    FlashVSR: Towards Real-Time Diffusion-Based Streaming Video Super-Resolution , author=. arXiv preprint arXiv:2510.12747 , year=

    work page arXiv

  18. [18]

    Stream-DiffVSR: Low-Latency Streamable Video Super-Resolution via Auto-Regressive Diffusion

    Stream-DiffVSR: Low-Latency Streamable Video Super-Resolution via Auto-Regressive Diffusion , author=. arXiv preprint arXiv:2512.23709 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    Venhancer: Generative space-time enhancement for video generation.arXiv preprint arXiv:2407.07667, 2024

    VEnhancer: Generative Space-Time Enhancement for Video Generation , author=. arXiv preprint arXiv:2407.07667 , year=

    work page arXiv

  20. [20]

    Longformer: The Long-Document Transformer

    Longformer: The Long-Document Transformer , author=. arXiv preprint arXiv:2004.05150 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 2004

  21. [21]

    Generating Long Sequences with Sparse Transformers

    Generating Long Sequences with Sparse Transformers , author=. arXiv preprint arXiv:1904.10509 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1904

  22. [22]

    International Conference on Learning Representations (ICLR) , year=

    Reformer: The Efficient Transformer , author=. International Conference on Learning Representations (ICLR) , year=

  23. [23]

    arXiv preprint arXiv:2502.18137 , year=

    SpargeAttention: Accurate and Training-free Sparse Attention Accelerating Any Model Inference , author=. arXiv preprint arXiv:2502.18137 , year=

    work page arXiv

  24. [24]

    Block-Sparse Attention , author=

  25. [25]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  26. [26]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  27. [27]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    VBench: Comprehensive Benchmark Suite for Video Generative Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  28. [28]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year=

    Scalable Diffusion Models with Transformers , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year=

  29. [29]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    High-Resolution Image Synthesis with Latent Diffusion Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  30. [30]

    Tiny Autoencoder for High-Resolution Video (

    Ollin Boer Bohan , howpublished=. Tiny Autoencoder for High-Resolution Video (

  31. [31]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Video Diffusion Models , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  32. [32]

    AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning

    AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning , author=. arXiv preprint arXiv:2307.04725 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets

    Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets , author=. arXiv preprint arXiv:2311.15127 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  34. [34]

    HaCohen, Yoav and Chiprut, Nisan and Brazowski, Benny and Shalem, Daniel and Moshe, Dudu and Richardson, Eitan and Levin, Eran and Shiran, Guy and Zabari, Nir and Gordon, Ori and Panet, Poriya and Weissbuch, Sapir and Kulikov, Victor and Bitterman, Yaki and Melumian, Zeev and Bibi, Ofir , journal=

  35. [35]

    Diffusion Models Are Real-Time Game Engines

    Diffusion Models Are Real-Time Game Engines , author=. arXiv preprint arXiv:2408.14837 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  36. [36]

    Pyramidal flow matching for efficient video generative modeling.arXiv preprint arXiv:2410.05954, 2024

    Pyramidal Flow Matching for Efficient Video Generative Modeling , author=. arXiv preprint arXiv:2410.05954 , year=

    work page arXiv

  37. [37]

    arXiv preprint arXiv:2402.09470 , year=

    Rolling Diffusion Models , author=. arXiv preprint arXiv:2402.09470 , year=

    work page arXiv

  38. [38]

    Kim, Jihwan and Kang, Junoh and Choi, Jinyoung and Han, Bohyung , journal=

  39. [39]

    arXiv preprint arXiv:2310.07702 , year=

    ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models , author=. arXiv preprint arXiv:2310.07702 , year=

    work page arXiv

  40. [40]

    arXiv preprint arXiv:2311.16973 , year=

    DemoFusion: Democratising High-Resolution Image Generation With No \ \ \ , author=. arXiv preprint arXiv:2311.16973 , year=

    work page arXiv

  41. [41]

    arXiv preprint arXiv:2403.02084 , year=

    ResAdapter: Domain Consistent Resolution Adapter for Diffusion Models , author=. arXiv preprint arXiv:2403.02084 , year=

    work page arXiv

  42. [42]

    arXiv preprint arXiv:2407.02158 , year=

    UltraPixel: Advancing Ultra-High-Resolution Image Synthesis to New Peaks , author=. arXiv preprint arXiv:2407.02158 , year=

    work page arXiv

  43. [43]

    Seedance 1.0: Exploring the Boundaries of Video Generation Models

    Seedance 1.0: Exploring the Boundaries of Video Generation Models , author=. arXiv preprint arXiv:2506.09113 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  44. [44]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  45. [45]

    Reward Forcing: Efficient Streaming Video Generation with Rewarded Distribution Matching Distillation

    Reward Forcing: Efficient Streaming Video Generation with Rewarded Distribution Matching Distillation , author=. arXiv preprint arXiv:2512.04678 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  46. [46]

    Self-Forcing++: Towards Minute-Scale High-Quality Video Generation

    Self-Forcing++: Towards Minute-Scale High-Quality Video Generation , author=. arXiv preprint arXiv:2510.02283 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  47. [47]

    arXiv preprint arXiv:2502.05179 , year=

    FlashVideo: Flowing Fidelity to Detail for Efficient High-Resolution Video Generation , author=. arXiv preprint arXiv:2502.05179 , year=

    work page arXiv

  48. [48]

    HunyuanVideo 1.5 Technical Report

    HunyuanVideo 1.5: A Systematic Framework For Large Video Generative Models , author=. arXiv preprint arXiv:2511.18870 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  49. [49]

    Fsvideo: Fast speed video diffusion model in a highly-compressed latent space.arXiv preprint arXiv:2602.02092, 2026

    FSVideo: Fast Speed Video Diffusion Model in a Highly-Compressed Latent Space , author=. arXiv preprint arXiv:2602.02092 , year=

    work page arXiv

  50. [50]

    Longcat-video technical report.arXiv preprint arXiv:2510.22200, 2025

    LongCat-Video Technical Report , author=. arXiv preprint arXiv:2510.22200 , year=

    work page arXiv

  51. [51]

    arXiv preprint arXiv:2506.13691 , year=

    UltraVideo: High-Quality UHD Video Dataset with Comprehensive Captions , author=. arXiv preprint arXiv:2506.13691 , year=

    work page arXiv

  52. [52]

    HaCohen, Yoav and Brazowski, Benny and Chiprut, Nisan and Bitterman, Yaki and Kvochko, Andrew and Berkowitz, Avishai and Shalem, Daniel and Lifschitz, Daphna and Moshe, Dudu and Porat, Eitan and Richardson, Eitan and Shiran, Guy and Chachy, Itay and Chetboun, Jonathan and Finkelson, Michael and Kupchick, Michael and Zabari, Nir and Guetta, Nitzan and Kotl...

  53. [53]

    Waver: Wave your way to lifelike video generation.arXiv preprint arXiv:2508.15761, 2025

    Waver: Wave Your Way to Lifelike Video Generation , author=. arXiv preprint arXiv:2508.15761 , year=

    work page arXiv

  54. [54]

    arXiv preprint arXiv:2603.21986 , year=

    Speed by Simplicity: A Single-Stream Architecture for Fast Audio-Video Generative Foundation Model , author=. arXiv preprint arXiv:2603.21986 , year=

    work page arXiv

  55. [55]

    Ke, Junjie and Wang, Qifei and Wang, Yilin and Milanfar, Peyman and Yang, Feng , booktitle=

  56. [56]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  57. [57]

    and Loy, Chen Change , booktitle=

    Wang, Jianyi and Chan, Kelvin C.K. and Loy, Chen Change , booktitle=. Exploring

  58. [58]

    2022 , eprint=

    LAION-5B: An Open Large-Scale Dataset for Training Next Generation Image-Text Models , author=. 2022 , eprint=

    2022

  59. [59]

    Scaling Laws of

    Liu, Xiaoran and Yan, Hang and Zhang, Shuo and An, Chenxin and Qiu, Xipeng and Lin, Dahua , journal=. Scaling Laws of

  60. [60]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Direct Preference Optimization: Your Language Model is Secretly a Reward Model , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  61. [61]

    Towards Accurate Generative Models of Video: A New Metric & Challenges

    FVD: A New Metric for Video Generation , author=. arXiv preprint arXiv:1812.01717 , year=

    work page internal anchor Pith review Pith/arXiv arXiv