Pith. sign in

REVIEW 3 major objections 5 minor 45 references

FreqForcing: Autoregressive Long Video Generation via Spectral Self-Anchoring

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read FreqForcing stops long-video drift by anchoring low-frequency content to the first frames, stretching a 5-second-trained model to stable two-minute generation without retraining.

desk verdict A genuinely novel training-free inference recipe for autoregressive video that mostly holds up, but the frozen-anchor assumption and same-set hyperparameter tuning keep it at conditional rather than accepted. read the letter →

arxiv 2607.27110 v2 pith:N4IJCKFV submitted 2026-07-29 cs.CV

Jiatong Li , Leo Liang , Linghe Kong , Yulun Zhang This is my paper
classification cs.CV
keywords autoregressivevideodiffusionlong-videogenerationspectralenergydriftfrequency-domainfusionattentionsinktraining-freeinferenceSelf-ForcingSelf-Anchoring
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 claims that the visual collapse seen in autoregressive long-video generation is, in the frequency domain, a drift of energy into low-frequency bands, and that a training-free inference-time correction can stop it. The proposed FreqForcing runs two attention branches — a local sliding window that preserves motion and detail, and an anchor branch fixed to the first high-quality frames — and fuses their outputs in the frequency domain, replacing the low-frequency part of the local output with the anchor's. On a Self-Forcing model trained on 5-second clips, this yields stable two-minute videos (24x extrapolation) with no retraining, matching or beating competitors on VBench-Long. The core insight is that the low-frequency spectrum of the first frames is a reliable reference for the whole rollout, so anchoring to it prevents drift without freezing the motion.

What carries the argument

Spectral Self-Anchoring (SSA), a dual-branch inference scheme with a local attention branch (sliding-window causal attention with attention sink) and an anchor attention branch (keys/values from a frozen cache of six early frames). The two outputs are fused in the frequency domain: after a 3D FFT, a spatial-temporal Gaussian low-pass filter H_lp weights a residual (A_anc − A_loc), so the fused output is A_loc + λH_lp(A_anc − A_loc). This frequency-selective gate pulls low-frequency content toward the anchor while leaving high-frequency motion untouched. The scheme is applied only in the first two denoising steps, exploiting the coarse-to-fine structure of diffusion.

What would settle it

Generate a 120-second video with a prompt that demands a major scene transition (e.g., a room gradually changing from day to night) using FreqForcing and compare the lighting and background at t=100s with a version that disables the anchor branch; if the anchored version keeps the original lighting and layout unchanged, the assumption that initial frames are valid anchors is falsified. Alternatively, measure the low-frequency spectral distance between the anchor frames and a ground-truth scene at each time step; when that distance grows, the method will force drift and produce repeated or froz

Watch

Extended reading notes

Core claim

The paper's central claim is that error accumulation in autoregressive video diffusion is primarily a low-frequency spectral energy drift, and that this drift can be corrected at inference time by Spectral Self-Anchoring (SSA). SSA maintains a frozen cache of six high-quality anchor frames from the first pretrained-horizon frames and, once generation exceeds the pretrained length, computes a second attention output against those anchors. The two outputs — local and anchor — are transformed with a 3D FFT, and a Gaussian low-pass filter is used to replace the low-frequency components of the local attention output with those of the anchor output via a weighted residual (Eq. 11). The paper provi

Load-bearing premise

The first six high-quality frames generated within the pretrained horizon remain a valid low-frequency reference for the entire two-minute rollout; if the scene legitimately evolves (lighting, layout, background), Eq. 11 will suppress that evolution and can push the video toward repetition.

Editorial extensions

If this is right

  • A model trained on 5-second clips can generate stable 120-second videos without any training, just by anchoring low-frequency attention outputs to the first frames.
  • Error accumulation in autoregressive video diffusion can be diagnosed as low-frequency spectral energy drift; attention sink alleviates but does not eliminate it, and SSA suppresses it further.
  • FreqForcing outperforms other training-free methods on VBench-Long dynamic degree and overall consistency, and is competitive with training-based methods.
  • The method adds only about 16.5% inference latency, making it practical for real-time streaming generation.

Reading between the lines

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

  • If the frequency-domain anchoring claim generalizes, other autoregressive modalities (audio, world models) could use a similar spectral anchor to prevent drift, provided a trustworthy early reference exists.
  • The method likely depends on the anchor frames remaining representative; for prompts with deliberate scene evolution (day-to-night, camera revealing new environments), the fixed low-frequency anchor may suppress legitimate changes, so an adaptive anchor-update rule is a natural testable extension.
  • The observation that attention sink only partially fixes drift while explicit spectral anchoring fixes it suggests that implicit context retention is weaker than explicit frequency-domain constraints for long-horizon stability.
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. This paper proposes FreqForcing, a training-free framework for long autoregressive video generation. The authors first observe that error accumulation during self-rollout manifests as low-frequency spectral energy drift in attention outputs and pixel frames. They then introduce Spectral Self-Anchoring (SSA), which runs a dual-branch attention inference: a local branch with an attention sink operates normally, while a frozen 'anchor' branch attends to high-quality frames from the early pretrained-horizon portion. The two attention outputs are fused in the frequency domain via Eq. (11), where the low-frequency component is blended with the anchor branch while the high-frequency component is retained from the local branch. SSA is applied only in the first two denoising steps and with a small anchor cache, adding about 16.5% latency. FreqForcing is evaluated on 60s and 120s VBench-Long generation using a Wan2.1-T2V-1.3B Self-Forcing base, reporting improved Dynamic Degree and Overall Consistency over training-free baselines and competitiveness with training-based methods. Ablations show the contribution of SSA and attention sink, and a hyperparameter study of lambda and sigma is included.

Significance. FreqForcing is a simple, training-free inference-time intervention that could be practically useful: it requires no retraining, the reported overhead is modest, and the authors provide a code link. The frequency-domain perspective on error accumulation is suggestive and connects to a growing literature on spectral guidance for video diffusion. If the VBench results survive a proper validation protocol, the method is a solid contribution: it improves Dynamic Degree and Overall Consistency over training-free baselines and is competitive with training-based methods on the tested prompts. However, the paper's spectral-preservation evidence is partly mechanical, because Eq. (11) forces low-frequency content toward the anchor branch by construction, and the frozen-anchor design has an unexamined failure mode for legitimate scene evolution. These issues should be addressed before the claims can be fully accepted.

major comments (3)
  1. [Table 2 and Section 4.2 (Evaluation)] The hyperparameters sigma and lambda are selected on the same 60s VBench-Long set used for the main comparison in Table 1, and no error bars or number of seeds are reported. With stochastic generation, the reported gaps between FreqForcing and training-free baselines (e.g., Dynamic Degree 59.58 vs 54.61; Overall Consistency 20.94 vs 20.84 at 60s) cannot be assessed for statistical significance. The chosen operating point (sigma=0.125, lambda=0.6) is also not a clear optimum in Table 2: larger sigma/lambda give higher Dynamic Degree and Overall Consistency at the cost of higher Repetitive Rate. I ask for either a held-out validation set for hyperparameter selection or multiple-seed means and standard deviations; without this, the 'outperforms existing training-free methods' claim is not yet established.
  2. [Section 3.4, Eq. (11); Fig. 6] The claim that SSA 'preserves spectral energy' is partly a built-in effect. Eq. (11) defines A_fused = A_loc + lambda*H_lp(A_anc - A_loc), so the low-frequency component of the fused attention is explicitly drawn toward the anchor branch. Consequently, Fig. 6's demonstration that low-frequency energy drift is reduced is largely a sanity check of the fusion formula rather than independent confirmation of the underlying mechanism. I recommend a control experiment, e.g., fusing a fixed random or deliberately low-quality anchor instead of the high-quality early frames, to show that the benefit is not solely the low-pass operation; alternatively, the paper should reframe Fig. 6 as a sanity check and rest the case on the VBench metrics.
  3. [Section 3.3-3.4, Eq. (11); Table 2] The frozen-anchor design may suppress legitimate scene evolution. The anchor cache is filled from roughly the first 18 latent frames (N_anc=6, one frame every three) and then frozen, and SSA is applied in the first two denoising steps where global layout and appearance are established. Eq. (11) then continuously biases the low-frequency content of all later chunks toward those early frames. For prompts whose correct continuation requires low-frequency changes--lighting shifts, camera moves, background transitions, day-to-night changes--this will actively pull the output back toward the initial scene. Table 2 already shows the expected footprint: Repetitive Rate rises monotonically with lambda (0.675 to 0.877). The paper acknowledges a consistency/diversity trade-off but never stress-tests prompts with major scene evolution. I request an evaluation on such prompts, reporting Dynamic Degre
minor comments (5)
  1. [Section 4.4, Table 2] The Repetitive Rate metric is introduced without a formal definition. Please specify the feature extraction, similarity threshold, and frame sampling used, so the ablation is reproducible.
  2. [Section 3.3 / 3.4] The anchor cache size N_anc=6 and the restriction of SSA to the first two denoising steps are design choices with a direct effect on both quality and latency, but they are not ablated. A short sensitivity study would strengthen the paper.
  3. [Section 3.1 / Fig. 3] The 'relative spectral energy' shown in Fig. 3 is not formally defined. Please specify the STFT window length, hop size, normalization, and whether the curves are averaged over layers, channels, or prompts.
  4. [Section 4.1 / Abstract] The relationship between 'pretrained on 5-second clips', L_pre=21, and the claimed 24x extrapolation to two minutes should be stated explicitly (frame rates, number of latent frames per second) so the extrapolation factor is unambiguous.
  5. [Section 4.3 / Fig. 4] The qualitative comparisons would be easier to assess if the supplementary material were summarized in the main text with a brief protocol for how video frames were selected and how many prompts were inspected.

Circularity Check

1 steps flagged · score 4.0 of 10

Spectral Self-Anchoring is defined to force low-frequency output toward the anchor branch, so Fig. 6's 'spectral preservation' is a built-in effect rather than an independent finding; external VBench results remain non-circular.

  1. self definitional [Section 'Spectral Anchoring', Eq. (11); Figure 6 caption]
    "ˆAfused = (1−λH lp) ˆAloc +λH lp ˆAanc = ˆAloc +λH lp( ˆAanc − ˆAloc) ... Figure 6: Effectiveness of SSA. SSA not only keeps the relative spectral energy closer to that of the initial frames, but also effectively suppresses its temporal drift."

    Eq. (11) defines A_fused so that its low-frequency content is a λ-weighted interpolation toward the anchor branch's low frequencies, and the anchor frames are the first pretrained-horizon frames frozen in the anchor cache. Therefore, the observation in Fig. 6 that SSA preserves spectral energy close to the initial frames is entailed by the construction itself, not an emergent empirical discovery. This is a mechanism-validation tautology rather than evidence that independently supports the method. The paper's headline VBench-Long comparisons and qualitative evaluations against baselines are external and not circular, so the circularity is partial.

full rationale

The core benchmarking claim—FreqForcing outperforms training-free baselines and remains competitive with training-based methods on VBench-Long—is supported by external metrics (Table 1) and qualitative comparisons, so the central result is not circular. The frequency-domain characterization of error accumulation (Fig. 3) is an empirical analysis of Self-Forcing and attention sink, and hyperparameter selection (Table 2) is based on external metrics. However, the paper's effectiveness evidence for Spectral Self-Anchoring in Fig. 6 reduces to the definition in Eq. (11): A_fused = A_loc + λH_lp(A_anc − A_loc) forces low-frequency energy toward the frozen anchor frames, so showing that spectral energy remains near the initial frames is a built-in effect. No load-bearing argument rests on self-citation, and no uniqueness theorem or ansatz is smuggled in via citation. The spectral-preservation claim is thus partially circular, but the independent benchmark evidence keeps the overall circularity score at 4 rather than higher.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on an empirical frequency-quality correlation, on the relevance of frozen early frames for the whole long horizon, on the sufficiency of early-step-only SSA, and on RoPE alignment. The method also depends on several hand-chosen free parameters (λ, σ, N_anc, S, step range, update interval). No new physical entity or conserved quantity is introduced; the anchor cache is a designed mechanism, not an independently falsifiable postulated entity.

free parameters (6)
  • λ (spectral anchoring strength) = 0.6
    Controls how much low-frequency content is pulled from anchor to local attention (Eq. 11). Chosen from Table 2 ablation balancing consistency against repetitive rate; not derived.
  • σ_xy = σ_t (low-pass filter bandwidth) = 0.125
    Gaussian low-pass width in spatial and temporal frequency (Eq. 10). Selected from the same Table 2 ablation on 60s VBench-Long generations; no held-out split reported.
  • N_anc (anchor cache capacity) = 6
    Number of frozen anchor latent frames; chosen to curtail overhead, with no independent justification.
  • S (attention sink size for local branch) = 12
    Taken from prior Deep Forcing / attention-sink practice; affects local-branch context stability.
  • SSA denoising-step range = first two of four steps (t4=1000, t3=937)
    Restricts SSA to early denoising based on a coarse-to-fine prior; an efficiency-versus-quality choice with no ablations over the exact step range.
  • anchor cache update interval = one frame per every 3 generated latent frames
    Design choice to fill the 6-slot cache within the first 18 frames; no ablation reported.
assumptions (5)
  • domain assumption STFT-based frequency-band energy of attention outputs and pixels is a faithful proxy for autoregressive error accumulation and visual quality.
    Used in Fig. 3 and the method motivation; no derivation or validation shows that spectral band energy maps one-to-one onto the failure modes (color drift, motion stagnation).
  • domain assumption Frozen anchor frames collected from early pretrained-horizon generations remain high-quality, relevant references for the entire two-minute rollout.
    Eq. 11 pulls low-frequency content of every future attention output toward these early frames; if scene content legitimately evolves, this assumption fails and causes repetition (see repetitive-rate rise in Table 2).
  • domain assumption Applying SSA only to the first two denoising steps (t4=1000, t3=937) is sufficient to stabilize the whole rollout because global layout is set early in denoising.
    The efficiency claim rests on the coarse-to-fine prior cited from Balaji et al. and Cao et al.; no experiment shows that late-step low-frequency drift is negligible.
  • domain assumption Rolling Forcing-style temporal RoPE alignment keeps relative token distances within the pretrained range without distorting attention.
    Adopted from prior work (Liu et al. 2025) and required for sink and anchor frames to remain in distribution when their temporal indices are shifted.
  • domain assumption Self-Forcing as the base model and VBench-Long metrics are valid external benchmarks for long-video quality.
    All comparisons build on Wan2.1 + Self-Forcing; VBench-Long scores are point estimates with no reported human evaluation or variance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FreqForcing: Autoregressive Long Video Generation via Spectral Self-Anchoring." pith.science (2026). https://pith.science/paper/N4IJCKFV

@misc{pith2026260727110,
  author       = {Pith},
  title        = {Pith review of: FreqForcing: Autoregressive Long Video Generation via Spectral Self-Anchoring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N4IJCKFV}},
  note         = {Machine review of arXiv:2607.27110}
}
read the original abstract

Autoregressive video diffusion models enable real-time streaming video generation. However, errors introduced during self-rollout accumulate over long horizons, manifesting as color drift, motion stagnation, and eventual visual collapse. In this paper, we characterize this phenomenon from a frequency-domain perspective: error accumulation appears as a pronounced energy drift in the low-frequency bands. We further investigate the effectiveness of attention sink in the frequency domain, and find that it improves the video quality by alleviating the spectral energy drift to some extent, but cannot fully resolve it. Motivated by the above analysis, we propose FreqForcing, a training-free framework that addresses error accumulation in long-video generation via Spectral Self-Anchoring (SSA). The proposed SSA leverages the low-frequency components of anchor attention to maintain long-horizon visual stability, while preserving dynamic motion through the high-frequency components of local attention. Our FreqForcing extends Self-Forcing pretrained on 5s clips to two-minute generation, achieving 24x extrapolation. Extensive experiments show that FreqForcing outperforms existing training-free methods quantitatively and qualitatively while remaining competitive with representative training-based approaches.

Figures

Figures reproduced from arXiv: 2607.27110 by the authors.

Figure 1
Figure 1. Comparison of different autoregressive video generation paradigms. The original Self-Forcing (Huang et al. 2025) suffers from severe error accumulation, while at￾tention sink can alleviate it to some extent. Our FreqForcing corrects this via Spectral Self-Anchoring (SSA). enables streaming generation also conditions each chunk on imperfect histories, so errors are repeatedly fed back and ac￾cumulate over long horizo… view at source ↗
Figure 2
Figure 2. Overview of FreqForcing. FreqForcing eliminates spectral energy drift during autoregressive video generation via Spectral Self-Anchoring (SSA). SSA consists of two steps. First, we design local-anchor attention branches. The local attention branch follows the standard sliding-window causal attention with attention sink to produce Aloc, while the anchor attention branch uses high-quality anchor frames to generate Aan… view at source ↗
Figure 3
Figure 3. Illustration of the relative spectral energy variations in both latent space and pixel space during 60s generation, including the DC component, low-frequency energy, and high-frequency energy. (a) The relative spectral energy of Self￾Forcing (Huang et al. 2025) collapses rapidly when attention sink is absent. (b) We investigate Self-Forcing with different attention sink sizes S ∈ {3, 12}. Increasing S alleviates the… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons on 120s videos. We compare FreqForcing with representative autoregressive video gener￾ation methods. Please refer to the supplementary material for more details. the absolute frame index n and sink set S = {1, 2, . . . , s}, the local historical…
Figure 5
Figure 5. Figure 5: Visual comparison for ablation. (a) Self-Forcing (Huang et al. 2025) suffers from severe color drift and visual fading. (b) Self-Forcing with attention sink size S = 12 alleviates color drift but still exhibits unsatisfactory results because of spectral drift. (c) Our …
Figure 6
Figure 6. Figure 6: Effectiveness of SSA. SSA not only keeps the relative spectral energy closer to that of the initial frames, but also effectively suppresses its temporal drift. First, we exploit the coarse-to-fine nature of the denoising process. Prior works (Balaji et al. 2022; Cao et…

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 16 linked inside Pith

  1. [1]

    FreeU : Free Lunch in Diffusion U-Net

    Si, Chenyang and Huang, Ziqi and Jiang, Yuming and Liu, Ziwei. FreeU : Free Lunch in Diffusion U-Net. CVPR

  2. [2]

    2024 , url=

    Video generation models as world simulators , author=. 2024 , url=

    2024

  3. [3]

    arXiv preprint arXiv:2511.18870 , year=

    Hunyuanvideo 1.5 technical report , author=. arXiv preprint arXiv:2511.18870 , year=

    arXiv

  4. [4]

    arXiv preprint arXiv:2503.20314 , year=

    Wan: Open and advanced large-scale video generative models , author=. arXiv preprint arXiv:2503.20314 , year=

    arXiv

  5. [5]

    ICML , year=

    Genie: Generative interactive environments , author=. ICML , year=

  6. [6]

    CVPR , year=

    Unisim: A neural closed-loop sensor simulator , author=. CVPR , year=

  7. [7]

    NeurIPS , year=

    Self forcing: Bridging the train-test gap in autoregressive video diffusion , author=. NeurIPS , year=

  8. [8]

    CVPR , year=

    From slow bidirectional to fast autoregressive video diffusion models , author=. CVPR , year=

Show all 45 references
  1. [9]

    arXiv preprint arXiv:2509.25161 , year=

    Rolling Forcing: Autoregressive Long Video Diffusion in Real Time , author=. arXiv preprint arXiv:2509.25161 , year=

    arXiv

  2. [10]

    arXiv preprint arXiv:2602.02214 , year=

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

    arXiv

  3. [11]

    NeurIPS , year=

    Freelong: Training-free long video generation with spectralblend temporal attention , author=. NeurIPS , year=

  4. [12]

    arXiv preprint arXiv:2507.00162 , year=

    Freelong++: Training-free long video generation via multi-band spectralfusion , author=. arXiv preprint arXiv:2507.00162 , year=

    arXiv

  5. [13]

    ICLR , year=

    Longlive: Real-time interactive long video generation , author=. ICLR , year=

  6. [14]

    CVPR , year=

    Infinity-rope: Action-controllable infinite video generation emerges from autoregressive self-rollout , author=. CVPR , year=

  7. [15]

    ICML , year=

    Deep forcing: Training-free long video generation with deep sink and participative compression , author=. ICML , year=

  8. [16]

    ICLR , year=

    Efficient streaming language models with attention sinks , author=. ICLR , year=

  9. [17]

    ICLR , year=

    Self-forcing++: Towards minute-scale high-quality video generation , author=. ICLR , year=

  10. [18]

    arXiv preprint arXiv:2602.07775 , year=

    Rolling sink: Bridging limited-horizon training and open-ended testing in autoregressive video diffusion , author=. arXiv preprint arXiv:2602.07775 , year=

    arXiv

  11. [19]

    arXiv preprint arXiv:2602.14027 , year=

    Train short, inference long: Training-free horizon extension for autoregressive video generation , author=. arXiv preprint arXiv:2602.14027 , year=

  12. [20]

    ICLR , year=

    Freenoise: Tuning-free longer video diffusion via noise rescheduling , author=. ICLR , year=

  13. [21]

    arXiv preprint arXiv:2605.06509 , year=

    FreeSpec: Training-Free Long Video Generation via Singular-Spectrum Reconstruction , author=. arXiv preprint arXiv:2605.06509 , year=

    arXiv

  14. [22]

    arXiv preprint arXiv:2503.00979 , year=

    Dialogue without limits: Constant-sized KV caches for extended responses in LLMs , author=. arXiv preprint arXiv:2503.00979 , year=

    arXiv

  15. [23]

    ICML , year=

    LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Models , author=. ICML , year=

  16. [24]

    arXiv preprint arXiv:2505.13211 , year=

    Magi-1: Autoregressive video generation at scale , author=. arXiv preprint arXiv:2505.13211 , year=

    arXiv

  17. [25]

    arXiv preprint arXiv:2504.13074 , year=

    Skyreels-v2: Infinite-length film generative model , author=. arXiv preprint arXiv:2504.13074 , year=

    arXiv

  18. [26]

    NeurIPS , year=

    Diffusion forcing: Next-token prediction meets full-sequence diffusion , author=. NeurIPS , year=

  19. [27]

    CVPR , year=

    Streamdit: Real-time streaming text-to-video generation , author=. CVPR , year=

  20. [28]

    arXiv preprint arXiv:2402.09470 , year=

    Rolling diffusion models , author=. arXiv preprint arXiv:2402.09470 , year=

    arXiv

  21. [29]

    ICLR , year=

    Autoregressive video generation without vector quantization , author=. ICLR , year=

  22. [30]

    ICLR , year=

    When attention sink emerges in language models: An empirical view , author=. ICLR , year=

  23. [31]

    ECCV , year=

    Freeinit: Bridging initialization gap in video diffusion models , author=. ECCV , year=

  24. [32]

    ICLR , year=

    FreqPrior: Improving Video Diffusion Models with Frequency Filtering Gaussian Noise , author=. ICLR , year=

  25. [33]

    IEEE Transactions on Acoustics, Speech, and Signal Processing , year=

    Short term spectral analysis, synthesis, and modification by discrete Fourier transform , author=. IEEE Transactions on Acoustics, Speech, and Signal Processing , year=

  26. [34]

    Huang, Ziqi and He, Yinan and Yu, Jiashuo and Zhang, Fan and Si, Chenyang and Jiang, Yuming and Zhang, Yuanhan and Wu, Tianxing and Jin, Qingyang and Chanpaisit, Nattapol and Wang, Yaohui and Chen, Xinyuan and Wang, Limin and Lin, Dahua and Qiao, Yu and Liu, Ziwei , booktitle=

  27. [35]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    arXiv

  28. [36]

    arXiv preprint arXiv:2410.13720 , year=

    Movie gen: A cast of media foundation models , author=. arXiv preprint arXiv:2410.13720 , year=

    arXiv

  29. [37]

    arXiv preprint arXiv:2211.01324 , year=

    ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers , author=. arXiv preprint arXiv:2211.01324 , year=

    arXiv

  30. [38]

    ICCV , year=

    Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing , author=. ICCV , year=

  31. [39]

    ICLR , year=

    Internvid: A large-scale video-text dataset for multimodal understanding and generation , author=. ICLR , year=

  32. [40]

    arXiv preprint arXiv:1404.1100 , year=

    A tutorial on principal component analysis , author=. arXiv preprint arXiv:1404.1100 , year=

    arXiv

  33. [41]

    NeurIPS , year=

    Attention is all you need , author=. NeurIPS , year=

  34. [42]

    ICCV , year=

    Scalable diffusion models with transformers , author=. ICCV , year=

  35. [43]

    Neurocomputing , year=

    Roformer: Enhanced transformer with rotary position embedding , author=. Neurocomputing , year=

  36. [44]

    arXiv preprint arXiv:2304.07193 , year=

    Dinov2: Learning robust visual features without supervision , author=. arXiv preprint arXiv:2304.07193 , year=

    arXiv

  37. [45]

    ICLR , year=

    Pyramidal Flow Matching for Efficient Video Generative Modeling , author=. ICLR , year=

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.