Pith. sign in

REVIEW 3 major objections 5 minor 6 cited by

TalkingMachines: Real-Time Audio-Driven FaceTime-Style Video via Autoregressive Diffusion Models

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims a pretrained video-diffusion transformer can be distilled into a two-step autoregressive model that streams audio-driven avatar video in real time with no error accumulation.

desk verdict A genuine engineering contribution to real-time talking-head streaming, but the signature claim of error-free infinite streaming is unsubstantiated. read the letter →

arxiv 2506.03099 v1 pith:AVTJUZDD submitted 2025-06-03 cs.SD cs.AIcs.GR

Chetwin Low , Weimin Wang This is my paper · ORCID
classification cs.SDcs.AIcs.GR
keywords audio-drivenvideogenerationautoregressivediffusiondistributionmatchingdistillationsparsecausalattentionreal-timestreamingtalkingheadanimationflow
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that a large pretrained video-diffusion model, which normally renders entire clips in one shot using bidirectional attention, can be turned into an autoregressive streamer that animates any still character image from live audio at interactive speed. Its central move is asymmetric distillation: a slow bidirectional teacher is compressed into a sparse causal student that emits short chunks of frames, with each chunk attending only to itself, the preceding chunk, and the first reference-image chunk. The authors claim this pattern eliminates the error accumulation that typically limits autoregressive video, and that with distillation down to two diffusion steps plus GPU-level scheduling the model generates frames faster than they replay. If correct, the approach gives a practical recipe for turning billion-parameter video foundation models into real-time conversational avatars that generalize to photorealistic, anime, and 3D characters rather than faces only.

What carries the argument

The load-bearing object is the sparse causal attention pattern. For a training clip of 81 frames (21 latent frames), the latent sequence is split into seven chunks of three frames, and every token in chunk $c_t$ attends only to tokens in its own chunk, in the previous chunk $c_{t-1}$, and in the starting chunk $c_0$ containing the ground-truth reference image; the paper argues this anchor to clean visual content is what blocks error accumulation while keeping the KV cache small. The second mechanism is asymmetric distillation: a bidirectional teacher, which sees all chunks at once, is distilled into the autoregressive student by distribution matching distillation, with the noise timestep kept global across chunks rather than sampled per chunk, because few-step generators follow a piecewise flow trajectory and should not attend to chunks at mismatched noise levels. A regression loss on the student's prediction stabilizes training alongside the distribution-matching objective.

What would settle it

Stream the model far beyond its training length — hundreds of chunks, both with continuous speech and with long silences — and measure identity consistency, color and contrast stability, and lip-sync scores as a function of distance from the reference chunk; if any quantity drifts monotonically as the stream grows, the no-error-accumulation claim fails. A sharper test: generate the same three-frame chunk given identical audio and previous chunk, but placed at stream positions 2 and 200; the outputs should be indistinguishable if the claim holds, and visibly divergent if it does not.

Watch

Extended reading notes

Core claim

The paper's claim, stated on its own terms, is that an 18-billion-parameter audio-conditioned video model—built by grafting audio cross-attention onto a pretrained 14-billion-parameter image-to-video diffusion transformer—can produce infinite-length, lip-synced talking video in real time, without error accumulation. The three load-bearing design choices are the sparse causal attention pattern just described, a global noise timestep shared by all chunks during distillation (so every chunk follows the same flow trajectory), and a two-step student trained by distribution matching distillation with a regression loss on mixed real and self-generated data. The authors report that the compute-cheapest configuration (three-frame chunks, two diffusion steps) needs only one H100 for the score model and one for the decoder, and that their disaggregated server keeps the time between generated chunks below the real-time threshold.

Load-bearing premise

The paper's central bet is that a model generating each short chunk of frames while looking only at the previous chunk and the very first reference image will keep generating coherent video forever, with no slow drift — and this bet is asserted rather than demonstrated by any long-horizon experiment.

Editorial extensions

If this is right

  • A single H100 for the score model plus one for the decoder suffices to generate three-frame chunks in real time, so the full 18-billion-parameter model becomes deployable on modest industrial hardware.
  • Because the model keeps a constant attention context (current, previous, and first chunk), generation cost per chunk stays constant as the stream grows, which is what makes arbitrarily long sessions practical.
  • The method inherits the visual generality of the pretrained foundation model rather than face-specific warping, extending real-time animation to full-body figures, anime, and stylized characters.
  • Switching a special silence embedding per character by turn of speech lets one model animate multiple characters in a single conversation, with each character attending to audio only when speaking.
  • Quality can be traded against hardware: the 7-frame-chunk, 4-step setting improves perceptual quality (FVD 208 vs 238) at the cost of four score-model GPUs, while lip-sync scores stay approximately constant across settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The no-error-accumulation claim rests on a local argument and an 81-frame clip design; the natural next experiment the paper does not report is a long-horizon stress test measuring identity drift, color drift, and lip-sync degradation over thousands of streamed frames.
  • The reference-chunk anchor $c_0$ is doing double duty as both identity memory and error correction; ablating it by dropping $c_0$ from later chunks would isolate how much of the stability comes from re-anchoring versus from the per-chunk training objective.
  • The speaking/silence embedding is a discrete on/off signal; a graded or voice-activity-derived attention gain could extend the same architecture to overlapping speech, music, or non-speech audio without retraining the video backbone.
  • The real-time criterion is defined as generation keeping pace with playback; a stricter measure—audio-to-first-visible-mouth-motion latency including the audio model's response time and network transport—is not reported, so 'real-time' as experienced by a user may include additional fixed overhead.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. TalkingMachines adapts the pretrained Wan2.1 14B image-to-video flow-matching DiT into an audio-driven avatar generator by adding a 1.2B audio projection module and audio cross-attention, then distills the bidirectional teacher into a two-step autoregressive student with sparse causal attention (each chunk attends to itself, the previous chunk, and the initial reference-image chunk). The paper also presents a disaggregated inference system with sequence-parallel diffusion workers and a dedicated VAE decoder, plus CUDA-stream overlap and KV caching. It claims infinite streaming without error accumulation and real-time latency suitable for interactive FaceTime-style applications.

Significance. If substantiated, the paper would be a significant practical contribution: it shows a concrete path from a large public video-generation model to an interactive streaming system, with meaningful engineering contributions (Score-VAE disaggregation, CUDA-stream overlap, KV caching) and a three-stage training recipe. The use of standard flow-matching and DMD objectives, the explicit sparse-causal attention design, and the system-level latency analysis are strengths. However, the central claims of no error accumulation, real-time operation, and high lip-sync quality rest on a single internal ablation with no baselines, no error bars, and no long-horizon evaluation, so the practical significance is not yet demonstrated.

major comments (3)
  1. [§4.2 / Abstract] The abstract and Section 4.2 claim that the sparse attention pattern 'prevents error accumulation, as the model always has access to the clean reference image in c0.' This inference is not warranted. The reference chunk anchors identity, but the autoregressive recurrence carries the generated dynamical state (pose, gaze, lip phase, motion continuity) through the generated chunk c_{t-1}, so a bias or drift in one chunk enters as conditioning for the next; a clean c0 does not by itself stop errors from compounding. The paper reports no long-horizon experiment: Table 1 measures FVD and SyncNet on fixed-length clips only, with no per-chunk-index breakdown, no comparison of early versus late chunks, and no baseline with full bidirectional context. A concrete remedy would be to report quality metrics as a function of chunk index on streams of at least several hundred chunks and to compare against a bidirectional-oracle baseline.
  2. [§5.4, Table 1] The central quality evaluation is a single 2x2 ablation on one internal dataset, without error bars over seeds or video sets. The text states that the degradation from FVD 208.2 to 238.1 is 'modest' and that lip-sync performance 'remains relatively consistent,' but with no variance estimates these statements are unsupported. Moreover, there is no comparison against the original Wan2.1 teacher at full 24-NFE bidirectional inference, no non-distilled autoregressive student, and no existing real-time talking-head baselines (e.g., VASA-1, MegaPortraits). This is particularly problematic because Section 4.2 asserts that the sparse causal attention is 'without any quality degradation,' which Table 1 directly contradicts: reducing chunk size from 7 to 3 raises FVD from 208.2 to 215.4 at 4 steps and from 235.1 to 238.1 at 2 steps. Please add baselines, multiple runs with error bars, and reconcile this contradiction.
  3. [§4.3, Figure 3] The real-time claim rests on the TTBC metric, but Figure 3 does not appear in the manuscript and no numeric TTBC values are reported. The definition of TTBC as 'time taken between chunks' is ambiguous: the paper does not state whether it is a mean, median, or tail latency, whether it is measured per chunk or per frame, or whether it includes VAE decoding, device-to-host transfer, and audio conditioning. Without these details and the plotted values, the statement that Cases 3 and 4 'consistently meet the real-time threshold' cannot be verified. Please provide the figure, a precise definition, the measured values, and the chunk duration that defines the real-time threshold.
minor comments (5)
  1. [§4.3] Figure 3 is referenced but the plot appears to be missing from the manuscript; please include it.
  2. [§4.2, Eq. (7)] Equation (7) is a generic attention formula and does not express the chunked sparse masking; write the mask explicitly (e.g., which positions are allowed for each query in chunk c_t).
  3. [§5.4, Table 1] Report the number of clips in the internal evaluation dataset and state whether Sync-C and Sync-D are averaged over frames or over clips; also indicate how many random seeds or runs are used for the FVD and SyncNet values.
  4. [§5.2] The training setup lists steps and GPU counts but not effective batch size or per-step time; adding these would improve reproducibility and help readers assess the compute cost.
  5. [References] Reference [10] is cited for the Rectified Flow formulation, but the citation points to 'Rectified flow matching: Towards fast and stable training of continuous normalizing flows' by Esser et al.; the original Rectified Flow paper is by Liu et al. (2022). Please verify and correct this citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the derivation rests on external pretrained weights, external distillation methods, and external evaluation metrics; no step reduces to its own inputs by construction.

full rationale

The paper's central technical chain adapts the pretrained WAN2.1 I2V model with external flow-matching and DMD objectives, and its sparse-causal autoregressive structure is an explicit modification of CausVid, an external prior work cited as [30]. No load-bearing step is justified by a self-citation: the references to WAN2.1, DMD, CausVid, MagicInfinite, Hallo3, SyncNet, and LiveKit are all external works with no author overlap with Chetwin Low and Weimin Wang. The DMD training does use student-generated synthetic samples, but this is a training-data augmentation technique, not a derivation of the evaluation outcome; the reported metrics (FVD, Sync-C, Sync-D) are computed on held-out clips using external measurement models. The 'infinite streaming without error accumulation' claim is asserted in Section 4.2 based on the clean reference chunk c0, and the paper provides no long-horizon evaluation to substantiate it; however, an unsupported empirical assertion is a correctness or evidence risk, not a circularity. Nothing in the paper fits a parameter to a target quantity and then renames that fit as a prediction, and no uniqueness or impossibility theorem is imported from the authors' own prior work. Overall, the derivation is not circular; the main weakness is missing evidence for the long-horizon claim, which is outside the circularity definition. Score 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The system rests on transferability of a pretrained video diffusion transformer and on the sufficiency of sparse causal attention for long streams; neither is established outside this paper's internal evidence.

free parameters (4)
  • Target resolution = 512x512
    Chosen as a lower resolution than the pretrained 480x832 to cut context length and latency (Section 5.1); affects all subsequent results.
  • Audio cross-attention window = 5 latent frames
    Local window-based alignment inspired by Hallo3 (Section 4.1); hand-set hyperparameter for temporally local audio-visual correlation.
  • Chunk size = 3 latent frames (versus 7)
    Selected from the 2x2 ablation on internal data (Section 5.4, Table 1) to balance quality against GPU cost.
  • Number of distillation steps = 2
    Selected from the ablation (Table 1) to trade a small FVD increase for a large reduction in required GPUs.
assumptions (4)
  • domain assumption The pretrained Wan2.1 image-to-video model can be adapted for audio-conditioned generation by adding audio cross-attention layers and fine-tuning.
    Used throughout Section 4.1 and Stages 1-2 of Section 5.2; the success of the method depends on this transferability.
  • domain assumption Speech-driven motion needs only temporally local audio, and each latent frame only needs to attend to current, previous, and starting chunks.
    Section 4.1 and 4.2 state this observation without derivation; it underwrites the sparse causal mask.
  • ad hoc to paper DMD asymmetric distillation from a bidirectional teacher to a sparse causal two-step student preserves distribution fidelity and prevents error accumulation.
    Section 4.2 relies on CausVid and DMD, but the specific claim that this prevents autoregressive error accumulation is not demonstrated with long-horizon metrics.
  • domain assumption Bidirectional attention in the teacher is the main latency bottleneck, and chunked causal generation with two steps and KV caching achieves real-time performance.
    Section 2 and Section 4.3; the TTBC evidence is qualitative in Figure 3, with no numerical table published in the text.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TalkingMachines: Real-Time Audio-Driven FaceTime-Style Video via Autoregressive Diffusion Models." pith.science (2026). https://pith.science/paper/AVTJUZDD

@misc{pith2026250603099,
  author       = {Pith},
  title        = {Pith review of: TalkingMachines: Real-Time Audio-Driven FaceTime-Style Video via Autoregressive Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AVTJUZDD}},
  note         = {Machine review of arXiv:2506.03099}
}
read the original abstract

In this paper, we present TalkingMachines -- an efficient framework that transforms pretrained video generation models into real-time, audio-driven character animators. TalkingMachines enables natural conversational experiences by integrating an audio large language model (LLM) with our video generation foundation model. Our primary contributions include: (1) We adapt a pretrained SOTA image-to-video DiT into an audio-driven avatar generation model of 18 billion parameters; (2) We enable infinite video streaming without error accumulation through asymmetric knowledge distillation from a bidirectional teacher model into a sparse causal, autoregressive student model; (3) We design a high-throughput, low-latency inference pipeline incorporating several key engineering optimizations such as: (a) disaggregation of the DiT and VAE decoder across separate devices, (b) efficient overlap of inter-device communication and computation using CUDA streams, (c) elimination of redundant recomputations to maximize frame-generation throughput. Please see demo videos here - https://aaxwaz.github.io/TalkingMachines/

Figures

Figures reproduced from arXiv: 2506.03099 by the authors.

Figure 1
Figure 1. TalkingMachines provides a framework to generate highly dynamic, immersive FaceTime [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the DMD training workflow for TalkingMachines. The diagram illustrates the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Runtime analysis comparing the latency of various server designs like a simple self [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. InteracVid: Building a Real Interactive Audio-Visual Response Dataset from Live-Chat Videos

    cs.CV 2026-08 conditional novelty 7.0 of 10

    InteracVid delivers 454K livestream-derived context-query-response triplets, pairing real or LLM-reconstructed chat triggers with real audio-video reactions, and shows fine-tuning gains on genuine queries.

  2. OmniMate: Open-Ended Real-Time Streaming Audio-Visual Generation for Interactive Avatars

    cs.CV 2026-07 conditional novelty 6.0 of 10

    OmniMate generates real-time streaming video, speech, and sound effects for interactive avatars while preserving the same face and voice across long multi-turn conversations.

  3. Avatar Forcing: Real-Time Interactive Head Avatar Generation for Natural Conversation

    cs.LG 2026-01 conditional novelty 6.0 of 10

    A causal diffusion-forcing model generates interactive head-avatar motion with 500ms motion-generation latency and learns expressive reactions via DPO with synthetic negative samples.

  4. Live Avatar: Streaming Real-time Audio-Driven Avatar Generation with Infinite Length

    cs.CV 2025-12 conditional novelty 6.0 of 10

    Live Avatar reports real-time streamable generation from a 14B audio-driven diffusion model at ~20 FPS on 5 H800s with stable identity over 10,000 seconds.

  5. SpeakerVid-5M: A Large-Scale High-Quality Dataset for Audio-Visual Dyadic Interactive Human Generation

    cs.CV 2025-07 conditional novelty 6.0 of 10

    SpeakerVid-5M provides 5.2 million audio-visual human clips (8,743 hours) with rich annotations and a dyadic interaction benchmark for training interactive virtual humans.

  6. MIDAS: Multimodal Interactive Digital-humAn Synthesis via Real-time Autoregressive Video Generation

    cs.CV 2025-08 reject novelty 5.0 of 10

    A new autoregressive video-generation framework for interactive digital humans, with a 64x compression autoencoder and a diffusion renderer, claims real-time multimodal control but is only demonstrated for audio input.

Reference graph

Works this paper leans on

31 extracted references · 12 canonical work pages · cited by 6 Pith papers

  1. [1]

    Paddleocr, awesome multilingual ocr toolkits based on paddlepaddle

    PaddlePaddle Authors. Paddleocr, awesome multilingual ocr toolkits based on paddlepaddle. https://github.com/PaddlePaddle/PaddleOCR, 2020

    2020

  2. [2]

    Video generation models as world simulators, 2024

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators, 2024. OpenAI Technical Report

    work page 2024

  3. [3]

    Pyscenedetect: Python and opencv-based scene cut/transition detection program & library

    Brandon Castellano. Pyscenedetect: Python and opencv-based scene cut/transition detection program & library

    work page

  4. [4]

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

    Boyuan Chen, Diego Martí Monsó, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion. Advances in Neural Information Processing Systems, 37:24081–24125, 2024

    2024

  5. [5]

    Out of time: automated lip sync in the wild

    Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision–ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13 , pages 251–263. Springer, 2017

    work page 2016

  6. [6]

    Hallo3: Highly dynamic and realistic portrait image animation with video diffusion transformer

    Jiahao Cui, Hui Li, Yun Zhan, Hanlin Shang, Kaihui Cheng, Yuqi Ma, Shan Mu, Hang Zhou, Jingdong Wang, and Siyu Zhu. Hallo3: Highly dynamic and realistic portrait image animation with video diffusion transformer. arXiv preprint arXiv:2412.00733, 2024

    arXiv 2024

  7. [7]

    FlashAttention-2: Faster attention with better parallelism and work partitioning

    Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. In International Conference on Learning Representations (ICLR) , 2024

    2024

  8. [8]

    Fu, Stefano Ermon, Atri Rudra, and Christopher Ré

    Tri Dao, Daniel Y . Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 10

    work page 2022

Show all 31 references
  1. [9]

    Megaportraits: One-shot megapixel neural head avatars, 2023

    Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars, 2023

    2023

  2. [10]

    Rectified flow matching: Towards fast and stable training of continuous normalizing flows

    Patrick Esser, Robin Rombach, Andreas Blattmann, and Björn Ommer. Rectified flow matching: Towards fast and stable training of continuous normalizing flows. arXiv preprint arXiv:2402.14851, 2024

    2024 arXiv

  3. [11]

    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 , 2022

    2022 arXiv

  4. [12]

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

    Sam Ade Jacobs, Masahiro Tanaka, Chengming Zhang, Minjia Zhang, Shuaiwen Leon Song, Samyam Rajbhandari, and Yuxiong He. Deepspeed ulysses: System optimizations for enabling training of extreme long sequence transformer models. arXiv preprint arXiv:2309.14509, 2023

    2023 arXiv

  5. [13]

    An introduction to variational autoencoders

    Diederik P Kingma, Max Welling, et al. An introduction to variational autoencoders. F ounda- tions and Trends® in Machine Learning , 12(4):307–392, 2019

    2019

  6. [14]

    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, 2024

    2024 arXiv

  7. [15]

    Livekit: Open-source webrtc infrastructure for real-time audio and video, 2025

    LiveKit. Livekit: Open-source webrtc infrastructure for real-time audio and video, 2025. Accessed: 2025-05-27

    2025

  8. [16]

    Scalable parallel programming with cuda: Is cuda the parallel programming model that application developers have been waiting for? Queue, 6(2):40–53, 2008

    John Nickolls, Ian Buck, Michael Garland, and Kevin Skadron. Scalable parallel programming with cuda: Is cuda the parallel programming model that application developers have been waiting for? Queue, 6(2):40–53, 2008

    2008

  9. [17]

    Zero: Memory optimiza- tions toward training trillion parameter models, 2020

    Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. Zero: Memory optimiza- tions toward training trillion parameter models, 2020

    2020

  10. [18]

    High-resolution image synthesis with latent diffusion models, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022

    2022

  11. [19]

    Laion-aesthetics: Predicting the aesthetic quality of images, 2022

    Christoph Schuhmann, Andreas Biewald, Aran Komatsuzaki, et al. Laion-aesthetics: Predicting the aesthetic quality of images, 2022. Accessed: 2025-05-27

    2022

  12. [20]

    Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024

    2024

  13. [21]

    Raft: Recurrent all-pairs field transforms for optical flow

    Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020

    2020

  14. [22]

    Emo: Emote portrait alive generating expressive portrait videos with audio2video diffusion model under weak conditions

    Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive generating expressive portrait videos with audio2video diffusion model under weak conditions. InEuropean Conference on Computer Vision, pages 244–260. Springer, 2024

    2024

  15. [23]

    Fvd: A new metric for video generation

    Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. InarXiv preprint arXiv:1812.01717, 2019

    2019 arXiv

  16. [24]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  17. [25]

    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, 2025

    2025 arXiv

  18. [26]

    Vasa-1: Lifelike audio-driven talking faces generated in real time, 2024

    Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time, 2024. 11

    2024

  19. [27]

    Magicinfinite: Generating infinite talking videos with your words and voice

    Hongwei Yi, Tian Ye, Shitong Shao, Xuancheng Yang, Jiantong Zhao, Hanzhong Guo, Terrance Wang, Qingyu Yin, Zeke Xie, Lei Zhu, et al. Magicinfinite: Generating infinite talking videos with your words and voice. arXiv preprint arXiv:2503.05978, 2025

    2025 arXiv

  20. [28]

    Improved distribution matching distillation for fast image synthesis

    Tianwei Yin, Michaël Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and Bill Freeman. Improved distribution matching distillation for fast image synthesis. Advances in neural information processing systems, 37:47455–47487, 2024

    2024

  21. [29]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 6613–6623, 2024

    2024

  22. [30]

    From slow bidirectional to fast causal video generators

    Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024

    2024

  23. [31]

    Pytorch fsdp: Experiences on scaling fully sharded data parallel, 2023

    Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien-Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, Alban Desmaison, Can Balioglu, Pritam Damania, Bernard Nguyen, Geeta Chauhan, Yuchen Hao, Ajit Mathews, and Shen Li. Pytorch fsdp: Experiences on sc...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.