pith. sign in

arxiv: 2605.25550 · v1 · pith:I24U2LDPnew · submitted 2026-05-25 · 💻 cs.DC

DisagFusion: Asynchronous Pipeline Parallelism and Elastic Scheduling for Disaggregated Diffusion Serving

Hantian Zha , Teng Ma , Yang Yong , Haiwen Fu , Ruiyang Ma , Wei Gao , Ruihao Gong , Xianglong Liu
show 2 more authors
Wei Wang Yunpeng Chai
This is my paper

Pith reviewed 2026-06-29 20:43 UTC · model grok-4.3

classification 💻 cs.DC
keywords disaggregated servingdiffusion modelspipeline parallelismelastic schedulingasynchronous executionGPU heterogeneitycontent generation pipelines
0
0 comments X

The pith

DisagFusion overlaps computation with communication via asynchronous pipelines and dynamically rebalances stage instances to serve disaggregated diffusion models at scale.

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

The paper addresses the memory and compute imbalance in diffusion models by splitting encoder, DiT, and decoder stages across separate heterogeneous GPUs rather than running them monolithically. It introduces asynchronous pipeline parallelism to hide stage handoff costs and a hybrid scheduler that predicts performance and applies runtime feedback to adjust instance counts as workloads shift. These mechanisms aim to cut pipeline bubbles, tolerate network jitter, and maintain efficiency without brittle static provisioning. A sympathetic reader would care because the approach promises to make large generative models runnable on mixed, lower-cost hardware while delivering substantially higher throughput and lower latency than unified deployments.

Core claim

DisagFusion introduces asynchronous pipeline parallelism that overlaps computation and stage-to-stage communication to reduce pipeline bubbles and mitigate network jitter, together with a hybrid instance scheduling strategy that combines lightweight performance prediction with runtime feedback to continuously rebalance instance ratios across stages under workload shifts, enabling throughput gains of 3.4x-20.5x and end-to-end latency reductions of 18.5x versus a monolithic baseline while supporting flexible deployment on heterogeneous GPUs.

What carries the argument

asynchronous pipeline parallelism combined with hybrid instance scheduling that uses performance prediction and runtime feedback to rebalance stage instance ratios

If this is right

  • Diffusion serving can be deployed cost-efficiently across mixed GPU types instead of requiring uniform high-end clusters.
  • Stage handoff overheads and network jitter no longer limit pipeline throughput once computation and communication are overlapped asynchronously.
  • Instance counts per stage can be adjusted on the fly without manual reprovisioning when request patterns change.
  • End-to-end latency drops by more than an order of magnitude because pipeline bubbles are minimized and stages run on hardware matched to their footprints.

Where Pith is reading between the lines

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

  • The same overlap and rebalancing ideas could apply to other multi-stage generative pipelines such as autoregressive video or audio models that also exhibit compute-memory imbalance.
  • Production clusters might shift from buying homogeneous GPU fleets toward cheaper heterogeneous pools if the scheduler reliably tracks changing ratios.
  • The approach suggests a general pattern for disaggregated serving of any model whose stages have mismatched resource demands, provided prediction models remain lightweight.

Load-bearing premise

Lightweight performance prediction combined with runtime feedback can continuously rebalance instance ratios across stages under fast-changing workloads without introducing new overheads that offset the reported gains.

What would settle it

Run the system on a workload that rapidly alternates between high-resolution and low-resolution prompts at varying batch sizes and measure whether the observed throughput and latency improvements hold or whether scheduling overheads erase them.

Figures

Figures reproduced from arXiv: 2605.25550 by Haiwen Fu, Hantian Zha, Ruihao Gong, Ruiyang Ma, Teng Ma, Wei Gao, Wei Wang, Xianglong Liu, Yang Yong, Yunpeng Chai.

Figure 1
Figure 1. Figure 1: Video Generation Pipeline based on Diffusion Transformers. 160 160.0 Emb Enc Trans Dec Wan Qwen Hunyuan 0 20 40 60 80 GB BF16 37.8 INT8 18.9 BF16 58.1 INT8 29.2 FP16 FP8 80.0 INT4 40.0 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Memory footprint of different models. (The red lines indicate GPU memory capacities: 24 GB for RTX 4090 and A10, 40 GB for A100, and 80 GB for A100 and H100.) the refined latents. Among these stages, the DiT module serves as the computational bottleneck, requiring extensive iterative processing to ensure output quality. 2.2 Workload Characteristics In this subsection, we analyze the model weights and com￾p… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of scalability for different deployment strategies (baseline and the disaggregated version). Heterogeneous compute intensity across stages. The computational complexity of generation is predominantly governed by the DiT. The DiT ❶ has a much larger parameter size (e.g., Wan 2.1-14B in BF16 has 28.0 GB of DiT weights), and ❷ its attention computation scales as 𝑂(𝑇 2 · 𝐷) for to￾kens𝑇 and hidden d… view at source ↗
Figure 5
Figure 5. Figure 5: Impact of network latency/jitter on synchronous inter-stage transfer. Here, “5%/0.2s” means that each transfer via the transfer engine has a 5% probability of incurring an additional 0.2-second delay. 2.4 Challenges and Opportunity By adopting a disaggregated deployment strategy, our design effectively circumvents the model loading bottleneck while incurring negligible overhead. However, this architectural… view at source ↗
Figure 6
Figure 6. Figure 6: Real-time throughput under varying request pa￾rameters. The first 15 minutes use 4-step distill requests, after 15 minutes, the requests switch to 1-step distill. Static161 denotes a static 1:6:1 instance ratio, Static152 is defined simi￾larly, and Dynamic denotes dynamic instance scheduling. memory management and prefill–decode splitting [11, 24, 43, 45]. However, these optimizations rely on assumptions t… view at source ↗
Figure 8
Figure 8. Figure 8: A request workflow between three stages. responsible for initial request dispatching, while other ser￾vices alternately transmit metadata to complete computation in their respective stages. The data plane consists of three computation stages (En￾code, Diffusion Transformer, and Decode) that process data through a decentralized pipeline. These stages exchange intermediate tensors via the mooncake transfer e… view at source ↗
Figure 9
Figure 9. Figure 9: Decentralized queue scheduling. a change is detected, the predictor 𝑔ˆ(·) estimates the tar￾get instance counts based on features from the recent trace. Following the allocation update, the loop skips subsequent reactive logic to prevent interference, advancing immediately to the next iteration. Lines 11–17 implement the scheduling strategy. We trigger a scale-out operation only when the service GPU utiliz… view at source ↗
Figure 10
Figure 10. Figure 10: Examples of images generated by DisagFusion [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: End-to-end latency comparison between LightX2V and DisagFusion in serving requests. same prompts and random seeds on a single node with 8 GPUs and 16 GPUs, respectively, and compare against the monolithic baseline under the same GPU budgets. We use a resolution of 832×480 and generate 81 frames for each request [PITH_FULL_IMAGE:figures/full_fig_p009_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of scalability for LightX2V and DisagFusion under different workloads. and the Qwen2512 experiments on eight RTX 4090 GPUs. As illustrated by the CDF curves in [PITH_FULL_IMAGE:figures/full_fig_p010_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Network latency comparison between DisagFu￾sion and DisagFusion’s synchronous variant. throughput on 4 and 8 GPUs. Similarly, for the I2V 4-step task (Figure 12b), DisagFusion scales from 2.34 to 10.5 QPM, surpassing the baseline by factors of 3.4× and 7.7×. In the evaluation using the Qwen2512 (4090) model (Figure 12c), DisagFusion successfully scales to 16 GPUs, achieving 43.7 QPM—more than double its 8… view at source ↗
Figure 15
Figure 15. Figure 15: Real-time throughput under varying request pa￾rameters on the H100 cluster. In the parameter-varying trace, we compare DisagFusion’s dynamic scheduling against fixed instance allocations (1:6:1 and 1:5:2). During the first 15 minutes (4-step requests), the DiT stage is the bottleneck for both fixed settings. The 1:6:1 allocation (DisagFusion-S161) achieves a throughput of 4.9 QPM, outperforming the 1:5:2 … view at source ↗
Figure 14
Figure 14. Figure 14: Real-time throughput performance under dy￾namic workloads. Under mild jitter, both synchronous and asynchronous de￾signs remain stable. However, as jitter becomes more severe, the synchronous baseline suffers drastic throughput drops— 22.5% under moderate jitter and 30.3% under severe jitter— because the upstream stage blocks until the downstream ac￾knowledges receipt, turning every network delay into GPU… view at source ↗
Figure 16
Figure 16. Figure 16: Comparison of GPU utilization and memory foot￾print under different deployment strategies (LightX2V and DisagFusion). request rate. Meanwhile, the 1:5:2 allocation (DisagFusion￾S152) suffers from a severe bottleneck in the first phase (4.75 QPM) and, despite recovering to 8.67 QPM later, still lags be￾hind DisagFusion. These results confirm that DisagFusion’s dynamic scheduling is essential for maximizing… view at source ↗
read the original abstract

Diffusion-based generation is increasingly powering production content pipelines; however, deploying these models at scale remains a significant challenge. Model weights frequently exceed the memory capacity of commodity GPUs, while the encoder, diffusion transformer (DiT), and decoder stages exhibit highly imbalanced computational and memory footprints. A natural remedy is disaggregated serving-running stages as separate services on heterogeneous GPUs-yet this introduces new bottlenecks, including stage handoff overheads and fast-changing workloads that make cross-stage provisioning and scheduling brittle. This paper presents DisagFusion, enabling asynchronous pipeline parallelism and elastic scheduling for disaggregated diffusion serving. First, DisagFusion introduces asynchronous pipeline parallelism that overlaps computation and stage-to-stage communication to reduce pipeline bubbles and mitigate network jitter. Second, DisagFusion employs a hybrid instance scheduling strategy that combines lightweight performance prediction with runtime feedback to continuously rebalance instance ratio across stages under workload shifts. We implement DisagFusion and evaluate it with modern diffusion models. Compared to a monolithic baseline, DisagFusion improves throughput by 3.4x-20.5x and reduces end-to-end latency by 18.5x, while enabling flexible, cost-efficient deployment across heterogeneous GPUs.

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 / 2 minor

Summary. The manuscript presents DisagFusion for disaggregated diffusion model serving. It proposes asynchronous pipeline parallelism to overlap computation with stage-to-stage communication and reduce bubbles/jitter, plus a hybrid scheduling strategy that combines lightweight performance prediction with runtime feedback to dynamically rebalance instance ratios across encoder/DiT/decoder stages under shifting workloads. The central empirical claim is that this yields 3.4x–20.5x higher throughput and 18.5x lower end-to-end latency versus a monolithic baseline while supporting cost-efficient heterogeneous GPU deployments.

Significance. If the performance numbers and negligible-overhead assumption hold under realistic traces, the work would be a meaningful practical contribution to scalable diffusion serving, directly addressing memory-capacity limits and stage imbalance that currently hinder production deployment. The hybrid prediction+feedback scheduler and async pipeline are concrete engineering advances that could influence disaggregated inference systems more broadly.

major comments (2)
  1. [Evaluation section (performance claims and scheduler description)] The headline speedups rest on the unquantified claim that the hybrid scheduling loop (lightweight prediction + runtime feedback for rebalancing) adds negligible overhead relative to the 3.4x–20.5x gains. No ablation, no measured rebalancing latency, and no breakdown of feedback cost versus pipeline-stall or bandwidth overhead appear in the evaluation; this is load-bearing for the central claim because rapid workload shifts could make the rebalancing loop itself the dominant cost.
  2. [Evaluation section] The abstract states that the strategy 'continuously rebalance[s] instance ratio across stages under workload shifts' but supplies no concrete workload traces, shift frequencies, or sensitivity analysis showing that the prediction model remains accurate enough to avoid conservative over-provisioning that would shrink the reported deltas.
minor comments (2)
  1. The abstract would be strengthened by naming the specific diffusion models, input resolutions, and heterogeneous GPU configurations used for the 3.4x–20.5x and 18.5x numbers.
  2. Notation for 'instance ratio' and the exact formulation of the lightweight performance predictor should be defined earlier and used consistently.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the evaluation section. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Evaluation section (performance claims and scheduler description)] The headline speedups rest on the unquantified claim that the hybrid scheduling loop (lightweight prediction + runtime feedback for rebalancing) adds negligible overhead relative to the 3.4x–20.5x gains. No ablation, no measured rebalancing latency, and no breakdown of feedback cost versus pipeline-stall or bandwidth overhead appear in the evaluation; this is load-bearing for the central claim because rapid workload shifts could make the rebalancing loop itself the dominant cost.

    Authors: We agree the manuscript lacks explicit ablations and measurements of rebalancing overhead. The design uses lightweight prediction to keep costs low, but without quantified data the claim is under-supported. In revision we will add an ablation isolating the feedback loop, measured rebalancing latencies under varying shift rates, and a breakdown versus pipeline and bandwidth costs. revision: yes

  2. Referee: [Evaluation section] The abstract states that the strategy 'continuously rebalance[s] instance ratio across stages under workload shifts' but supplies no concrete workload traces, shift frequencies, or sensitivity analysis showing that the prediction model remains accurate enough to avoid conservative over-provisioning that would shrink the reported deltas.

    Authors: The current evaluation demonstrates rebalancing under shifts but does not include the requested traces, frequencies, or sensitivity results. We will add representative workload traces, quantified shift frequencies, and sensitivity analysis of the prediction model in the revised evaluation to substantiate accuracy and rule out over-provisioning effects. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on empirical implementation and evaluation

full rationale

The paper presents a systems design for disaggregated diffusion serving using asynchronous pipeline parallelism and a hybrid scheduling strategy (lightweight prediction plus runtime feedback). All headline performance numbers (3.4x-20.5x throughput, 18.5x latency) are stated as results of implementing the system and measuring it against a monolithic baseline. No equations, fitted parameters, self-citations, or uniqueness theorems appear in the provided text; the scheduling approach is described at the architectural level without any reduction of outputs to inputs by construction. The derivation chain is therefore self-contained empirical measurement rather than tautological.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are identifiable from the abstract alone.

pith-pipeline@v0.9.1-grok · 5764 in / 965 out tokens · 24337 ms · 2026-06-29T20:43:44.700541+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

48 extracted references · 19 canonical work pages · 12 internal anchors

  1. [1]

    Gulavani, Alexey Tumanov, and Ramachandran Ramjee

    Amey Agrawal, Nitin Kedia, Ashish Panwar, Jayashree Mohan, Nipun Kwatra, Bhargav S. Gulavani, Alexey Tumanov, and Ramachandran Ramjee. 2024. Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve. arXiv preprint arXiv:2403.02310.https://doi.org/ 10.48550/arXiv.2403.02310

    work page doi:10.48550/arxiv.2403.02310 2024

  2. [2]

    Andreas Blattmann et al . 2023. VideoLDM: Latent Video Diffu- sion Models for High-Fidelity Video Generation.arXiv preprint arXiv:2304.08818(2023).https://doi.org/10.48550/arXiv.2304.08818

    work page doi:10.48550/arxiv.2304.08818 2023

  3. [3]

    Language Models are Few-Shot Learners

    Tom B. Brown et al. 2020. Language Models are Few-Shot Learners. arXiv preprint arXiv:2005.14165(2020).https://doi.org/10.48550/arXiv .2005.14165

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv 2020

  4. [4]

    Patrick H Coppock, Brian Zhang, Eliot H Solomon, Vasilis Kypriotis, Leon Yang, Bikash Sharma, Dan Schatzberg, Todd C Mowry, and Dimitrios Skarlatos. 2025. LithOS: An operating system for efficient machine learning on GPUs. InProceedings of the ACM SIGOPS 31st Symposium on Operating Systems Principles. 1–17

    2025

  5. [5]

    Daniel Crankshaw et al. 2020. InferLine: Latency-aware Provisioning and Scaling for Prediction Serving Pipelines. InProceedings of the 2020 ACM Symposium on Operating Systems Principles (SOSP 20)

    2020

  6. [6]

    Yuwei Guo et al. 2023. AnimateDiff: Animate Your Personalized Text- to-Image Diffusion Models without Specific Tuning.arXiv preprint arXiv:2307.04725(2023).https://doi.org/10.48550/arXiv.2307.04725

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2307.04725 2023

  7. [7]

    Jonathan Ho et al. 2022. Imagen Video: High Definition Video Gener- ation with Diffusion Models.arXiv preprint arXiv:2210.02303(2022). https://doi.org/10.48550/arXiv.2210.02303

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2210.02303 2022

  8. [8]

    Jonathan Ho, William Chan, Chitwan Saharia, et al . 2022. Video Diffusion Models.arXiv preprint arXiv:2204.03458(2022).https: //doi.org/10.48550/arXiv.2204.03458

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2204.03458 2022

  9. [9]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models.arXiv preprint arXiv:2006.11239(2020).https: //doi.org/10.48550/arXiv.2006.11239

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2006.11239 2020

  10. [10]

    Xiaoxiao Jiang, Suyi Li, Lingyun Yang, Tianyu Feng, Zhipeng Di, Weiyi Lu, Guoxuan Zhu, Xiu Lin, Kan Liu, Yinghao Yu, et al. 2026. FlashPS: Efficient Generative Image Editing with Mask-aware Caching and Scheduling. InProceedings of the 21st European Conference on Computer Systems. 2109–2125

    2026

  11. [11]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica

  12. [12]

    Efficient Memory Management for Large Language Model Serving with PagedAttention

    Efficient Memory Management for Large Language Model Serving with PagedAttention. arXiv preprint arXiv:2309.06180.https: //doi.org/10.48550/arXiv.2309.06180

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2309.06180

  13. [13]

    Suyi Li, Lingyun Yang, Xiaoxiao Jiang, Hanfeng Lu, Dakai An, Zhipeng Di, Weiyi Lu, Jiawei Chen, Kan Liu, Yinghao Yu, et al . 2025. Katz: Efficient workflow serving for diffusion models with many adapters. In2025 USENIX Annual Technical Conference (USENIX ATC 25). 1037– 1052

    2025

  14. [14]

    Yifan Li, Zhaoyang Zhang, Haoyang Wu, Zhenhua Zheng, Hao Zhang, and Kaiming Chen. 2024. SwiftDiffusion: Efficient Diffusion Model Serving with Add-on Modules.arXiv preprint arXiv:2407.02031(2024). https://doi.org/10.48550/arXiv.2407.02031

    work page doi:10.48550/arxiv.2407.02031 2024

  15. [15]

    Zhen Li, Chen Feng, Yuhang Yang, Zongyi Wang, Yuxuan Zhang, and Wei Chen. 2024. DistriFusion: Distributed Parallel Inference for High-Resolution Diffusion Models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2024

  16. [16]

    Gonzalez, and Ion Stoica

    Zhuohan Li, Lianmin Zheng, Yinmin Zhong, Vincent Liu, Ying Sheng, Xin Jin, Yanping Huang, Zhifeng Chen, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. AlpaServe: Statistical Multiplexing with Model Parallelism for Deep Learning Serving. In17th USENIX Symposium on Operating Systems Design and Implementation (OSDI 23). USENIX Association, 663–679

    2023

  17. [17]

    Yanying Lin, Shuaipeng Wu, Shutian Luo, Hong Xu, Haiying Shen, Chong Ma, Min Shen, Le Chen, Chengzhong Xu, Lin Qu, et al. 2025. Understanding Diffusion Model Serving in Production: A Top-Down Analysis of Workload, Scheduling, and Resource Efficiency. InPro- ceedings of the 2025 ACM Symposium on Cloud Computing. 1–15

    2025

  18. [18]

    Yuan Liu, Jinyang Zhang, Ming Xu, Wei Li, Kai Chen, and Hao Zhang

  19. [19]

    LegoDiffusion: Micro-Serving Text-to-Image Diffusion Workflows

    LegoDiffusion: Modular Diffusion Models with Pluginable Adapters. arXiv preprint arXiv:2604.08123.https://doi.org/10.4 8550/arXiv.2604.08123

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    Michael Luo, Aaron Hao, Zhengxu Yan, Chengkun Cao, and Quang Luong Nhat Nguyen. 2025. DiT-Serve: An Efficient Serving Engine for Diffusion Transformers. arXiv preprint. To appear

    2025

  21. [21]

    ModelTC. 2025. LightX2V. GitHub repository.https://github.com/M odelTC/LightX2V

    2025

  22. [22]

    Mooncake Project. 2025. Mooncake Transfer Engine Design. Online documentation.https://kvcache-ai.github.io/mooncake/docs/transfe r-engine.html

    2025

  23. [23]

    Deepak Narayanan et al. 2020. Clockwork: Predictable Performance for Unpredictable Workloads. In14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20)

    2020

  24. [24]

    NVIDIA. 2024. GPUDirect RDMA (NVIDIA Documentation). NVIDIA Developer Documentation.https://docs.nvidia.com/cuda/gpudirect- rdma/

    2024

  25. [25]

    William Peebles and Saining Xie. 2023. Scalable Diffusion Models with Transformers. arXiv preprint arXiv:2212.09748.https://doi.org/ 10.48550/arXiv.2212.09748

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2212.09748 2023

  26. [26]

    Ruoyu Qin, Zheming Li, Weiran He, Jialei Cui, Heyi Tang, Feng Ren, Teng Ma, Shangming Cai, Yineng Zhang, Mingxing Zhang, et al. 2024. Mooncake: A kvcache-centric disaggregated architecture for llm serv- ing.ACM Transactions on Storage(2024)

    2024

  27. [27]

    Haoran Qiu, Anish Biswas, Zihan Zhao, Jayashree Mohan, Alind Khare, Esha Choukse, Íñigo Goiri, Zeyu Zhang, Haiying Shen, Chetan Bansal, et al. 2025. Modserve: Scalable and resource-efficient large multimodal model serving.arXiv preprint arXiv:2502.00937(2025)

    work page arXiv 2025

  28. [28]

    Qwen Team. 2024. Qwen-Image-2512 (Hugging Face model card). Hugging Face.https://huggingface.co/Qwen/Qwen-Image-2512

    2024

  29. [29]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with La- tent Diffusion Models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 10684–10695

    2022

  30. [30]

    Chitwan Saharia, William Chan, Saurabh Saxena, et al. 2022. Photore- alistic Text-to-Image Diffusion Models with Deep Language Under- standing.arXiv preprint arXiv:2205.11487(2022).https://doi.org/10.4 8550/arXiv.2205.11487

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  31. [31]

    SGLang Project. 2025. SGLang Diffusion Documentation. Online documentation.https://docs.sglang.ai/en/latest/diffusion/

    2025

  32. [32]

    SGLang Team. 2025. SGLang Diffusion: Serving Diffusion Models with SGLang. Blog post.https://lmsys.org/blog/2025-04-21-sglang- diffusion/ 13

    2025

  33. [33]

    Fu, Zhiqiang Xie, Beidi Chen, Clark Barrett, Joseph E

    Ying Sheng, Lianmin Zheng, Binhang Yuan, Zhuohan Li, Max Ryabinin, Daniel Y. Fu, Zhiqiang Xie, Beidi Chen, Clark Barrett, Joseph E. Gonzalez, Percy Liang, Christopher Ré, Ion Stoica, and Ce Zhang. 2023. FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU.arXiv preprint arXiv:2303.06865(2023).https://doi.org/10.48550/ar...

    work page doi:10.48550/arxiv.2303.06865 2023

  34. [34]

    Make-A-Video: Text-to-Video Generation without Text-Video Data

    Uriel Singer, Adam Polyak, J. Zohar, et al . 2022. Make-A-Video: Text-to-Video Generation without Text-Video Data.arXiv preprint arXiv:2209.14792(2022).https://doi.org/10.48550/arXiv.2209.14792

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2209.14792 2022

  35. [35]

    SoPrompts. 2026. Sora vs Runway vs Pika: Comparison. Blog post. https://soprompts.com/blog/sora-vs-runway-vs-pika

    2026

  36. [36]

    Stability AI et al. 2023. Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets.arXiv preprint arXiv:2311.15127 (2023).https://doi.org/10.48550/arXiv.2311.15127

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2311.15127 2023

  37. [37]

    Tencent. 2024. HunyuanVideo (GitHub repository). GitHub.https: //github.com/Tencent/HunyuanVideo

    2024

  38. [38]

    Vidwave. 2026. Pika Labs vs Stable Diffusion Video: Quality Test Results. Blog post.https://vidwave.ai/pika-labs-vs-stable-diffusion- video-quality-test-results

    2026

  39. [39]

    Ruben Villegas et al . 2022. Phenaki: Variable Length Video Gen- eration from Open Domain Textual Descriptions.arXiv preprint arXiv:2210.02399(2022).https://doi.org/10.48550/arXiv.2210.02399

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2210.02399 2022

  40. [40]

    Wan-AI. 2024. Wan2.1-T2V-14B (Hugging Face model card). Hugging Face.https://huggingface.co/Wan-AI/Wan2.1-T2V-14B

    2024

  41. [41]

    Wan-Video. 2025. Wan2.2 (GitHub repository). GitHub.https: //github.com/Wan-Video/Wan2.2

    2025

  42. [42]

    Ziyu Wang, Zhe Li, Yao Xu, Yuxuan Zhang, Le Chen, and Hao Zhang

  43. [43]

    DiffServe: Efficiently Serving Text-to-Image Diffusion Models with Query-Aware Model Scaling.arXiv preprint arXiv:2411.15381 (2024).https://doi.org/10.48550/arXiv.2411.15381

    work page doi:10.48550/arxiv.2411.15381 2024

  44. [44]

    Jay Zhangjie Wu, Yixiao Ge, et al . 2023. Tune-A-Video: One-Shot Tuning of Image Diffusion Models for Text-to-Video Generation.arXiv preprint arXiv:2212.11565(2023).https://doi.org/10.48550/arXiv.2212. 11565

    work page doi:10.48550/arxiv.2212 2023

  45. [45]

    Zhewei Yao et al. 2022. DeepSpeed Inference: Enabling Efficient Infer- ence of Transformer Models at Scale. InProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis (SC)

    2022

  46. [46]

    Gyeong-In Yu, Joo Seong Jeong, Geon-Woo Kim, Soojeong Kim, and Byung-Gon Chun. 2022. Orca: A Distributed Serving System for Transformer-Based Generative Models. In16th USENIX Symposium on Operating Systems Design and Implementation (OSDI 22). USENIX Association, 521–538

    2022

  47. [47]

    ZeroMQ Community. 2026. ZeroMQ. Project website.https://zero mq.org/

    2026

  48. [48]

    Yinmin Zhong, Shengyu Liu, Junda Chen, Jianbo Hu, Yibo Zhu, Xu- anzhe Liu, Xin Jin, and Hao Zhang. 2024. DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving. In18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24). USENIX Association, 193–210. 14

    2024