Pith. sign in

REVIEW 4 major objections 7 minor 45 references

Error buildup in long autoregressive video is low-frequency energy drift, and freezing early frames as spectral anchors corrects it without retraining.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 11:28 UTC pith:N4IJCKFV

load-bearing objection Solid training-free AR video fix with a clean frequency diagnosis; the missing control is whether early-frame anchors beat generic low-pass smoothing, but the paper still earns referee time. the 4 major comments →

arxiv 2607.27110 v1 pith:N4IJCKFV submitted 2026-07-29 cs.CV

FreqForcing: Autoregressive Long Video Generation via Spectral Self-Anchoring

Jiatong Li , Leo Liang , Linghe Kong , Yulun Zhang This is my paper
classification cs.CV
keywords autoregressive video diffusionlong video generationspectral energy driftSpectral Self-Anchoringattention sinktraining-free extrapolationfrequency-domain fusion
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Autoregressive video diffusion can stream frames in real time, but feeding its own imperfect history back into the model makes quality collapse over long horizons—color drift, frozen motion, then visual failure. This paper shows that collapse shows up clearly in the frequency domain as energy drifting in the DC and low-frequency bands that carry color, layout, and identity, while high-frequency detail becomes unstable. Attention sinks slow that drift but do not stop it. FreqForcing fixes the trajectory at inference time with Spectral Self-Anchoring: it runs a local attention branch for recent motion and an anchor branch on a small frozen cache of early high-quality frames, then fuses them so low frequencies come from the anchors and high frequencies stay local. Applied to a model trained only on 5-second clips, the method reaches stable two-minute generation (24× longer) and beats other training-free long-video tricks on dynamics and consistency while staying competitive with methods that retrain.

Core claim

Error accumulation during autoregressive self-rollout is spectral energy drift in the low-frequency bands; Spectral Self-Anchoring corrects that drift by injecting low-frequency components from a frozen cache of early high-quality frames into local attention, which extends a 5-second Self-Forcing model to stable two-minute video without any extra training.

What carries the argument

Spectral Self-Anchoring (SSA): dual local/anchor attention branches whose outputs are fused in the 3D frequency domain with a Gaussian low-pass gate so low frequencies track the anchor cache and high frequencies keep local motion (Eqs. 7–12), applied only in the first two denoising steps.

Load-bearing premise

A small frozen set of early high-quality frames remains a valid global low-frequency anchor for the whole long rollout when mixed in only at the first two denoising steps with fixed filter settings.

What would settle it

Generate minute-scale videos where early frames are a poor match for later required layout or identity (scene change, new objects, strong lighting shift) and check whether low-frequency energy still stabilizes and VBench-Long dynamic degree and overall consistency hold, or whether drift and collapse return despite SSA.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Training-free 24× horizon extension becomes practical for streaming autoregressive video diffusion trained only on short clips.
  • Attention-sink KV tricks are incomplete; explicit low-frequency anchoring of attention outputs is needed for minute-scale stability.
  • Restricting spectral correction to early denoising steps keeps overhead modest (~16.5% latency) while still locking global layout.
  • Dynamic degree and overall consistency on 60s/120s VBench-Long can match or beat several training-based long-video systems without distillation.

Where Pith is reading between the lines

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

  • The same low-frequency residual injection pattern may transfer to other causal rolling-cache generators (audio, 3D, world models) where DC drift appears before high-frequency failure.
  • If anchor frames are chosen or refreshed by content change rather than a fixed early window, SSA could handle shot cuts and interactive control without retraining.
  • Failure modes when the frozen anchor disagrees with the prompt’s later semantics would cleanly separate “spectral stabilization” from “semantic controllability.”

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The manuscript studies error accumulation in autoregressive video diffusion (Self-Forcing) from a frequency-domain perspective, arguing that rollout drift manifests as energy drift in the DC/low-frequency bands of attention outputs and pixel frames, and that attention sinks only partially suppress it. It then proposes FreqForcing, a training-free inference intervention: a dual-branch attention scheme in which a frozen cache of N_anc=6 early frames (gathered within the 5 s pretrained horizon) supplies an "anchor" attention output, whose low-frequency components are fused into the local sliding-window attention output via a Gaussian spectral gate (σ=0.125, λ=0.6) applied only in the first two of four denoising steps. The authors report stable 60 s and 120 s generation (24× extrapolation), improvements on VBench-Long dynamic degree and overall consistency over training-free baselines (Infinity-RoPE, Deep Forcing), competitiveness with training-based methods (LongLive, Rolling Forcing), ~16.5% latency overhead, and ablations over σ/λ (including a repetition-rate metric) and over components.

Significance. If the central claim holds, this is a practically useful result: minute-scale, real-time-compatible long-video generation from a 5 s-pretrained model with no retraining and modest latency cost, on open base models (Wan2.1/Self-Forcing), which makes it reproducible in principle. The frequency-domain characterization of error accumulation (Fig. 3), including the analysis of attention-sink size and of Deep Forcing's spectral jitter, is an independently valuable diagnostic contribution. The evaluation covers both 60 s and 120 s horizons with component and hyperparameter ablations and a repetition-rate metric, which is more than typical for training-free inference interventions. The main limitations on significance are the single-seed 128-prompt evaluation and the absence of controls isolating the anchoring mechanism from generic temporal smoothing, both noted above.

major comments (4)
  1. [§Method / Tab. 3] The central mechanistic claim is that anchoring to *early, high-quality* frames (Eq. 8: the anchor cache is filled only within the pretrained horizon L_pre=21 and then frozen) is what stabilizes the rollout. However, the component ablation in Tab. 3 only removes SSA or the attention sink entirely. There is no control that isolates provenance: (a) an anchor cache refreshed from recent rollout frames, or (b) low-frequency components extracted from the local branch itself (pure temporal low-pass of A_loc). If either performs comparably, the active ingredient is generic low-pass smoothing plus a larger effective sink, not 'self-anchoring' to early quality, and the paper's framing and its implicit robustness story would need revision. These controls are cheap (inference-only) and are the single most informative missing experiment. Relatedly, no experiment probes the failure mode where early-h
  2. [Tab. 1 vs. Tab. 2 (Experiments)] The paper's own Tab. 2 shows repetitive rate rising monotonically with λ and σ (0.675 → 0.877), i.e., the mechanism trades diversity for consistency, and the chosen operating point (σ=0.125, λ=0.6) is a hand-set point on that trade-off curve. Yet Tab. 1 reports no repetition/diversity metric for any method (Deep Forcing's 0.812 is mentioned only in passing), while the headline metrics where FreqForcing wins — Overall Consistency, Subject/Background Consistency — reward exactly the temporal similarity SSA injects. At 60 s, FreqForcing's Overall Consistency lead over LongLive is 20.94 vs 20.89, and Infinity-RoPE beats it on Dynamic Degree in some ablation rows of Tab. 2 (e.g., σ=0.5, λ=0.6 gives 62.68). The claim 'outperforms existing training-free methods' (Abstract, Tab. 1) is therefore evaluated on a metric suite tilted toward the method's own intervention. Please add repetitive rate (
  3. [§Experiments / Tab. 1] All quantitative results (Tabs. 1–3) are single-run numbers on the first 128 MovieGen prompts, with no seeds, error bars, or per-prompt variance. Several reported margins are small relative to plausible prompt-level variance (e.g., Overall Consistency 20.94 vs 20.89 at 60 s; Background Consistency differences in the third decimal). Meanwhile the free parameters (σ, λ, N_anc=6, S=12, SSA restricted to the first two denoising steps) are all selected by ablation on the same task family and the same 60 s setting used for the main table. Multi-seed evaluation (or at minimum bootstrap confidence intervals over prompts) and confirmation that hyperparameters were not selected on the reported test prompts are needed to make the Tab. 1 comparisons meaningful.
  4. [§Method, 'Local-anchor Attention Branches' (Eq. 8)] The frozen-anchor design assumes early frames remain a valid global low-frequency reference for the full 120 s, but nothing in the paper characterizes when this fails: prompts with legitimate global scene changes, or cases where the first 21 latent frames are themselves imperfect. The method section asserts the cache 'collects high-quality frames' without defining or verifying quality. A short failure-mode analysis (e.g., correlation between early-frame quality and final VBench scores, or examples where over-anchoring freezes content that should evolve) would substantiate the 24×-extrapolation claim beyond the average case and clarify the scope of the method.
minor comments (7)
  1. [Fig. 3, Fig. 6] Fig. 3 and Fig. 6: 'relative spectral energy' is never defined (relative to what baseline?), and the STFT protocol (window length, hop, band boundaries for DC/low/high) is not specified. These plots carry much of the diagnostic argument, so the measurement details should be given.
  2. [Eq. 8] Eq. 8: the case A_anc = 0 for L_gen ≤ L_pre is clear in isolation, but the text should state explicitly that spectral anchoring (Eqs. 9–12) is skipped entirely in this regime, since Eq. 11 with A_anc = 0 would otherwise subtract λH_lp·A_loc.
  3. [§Method] Anchor cache filling: 'for every three generated latent frames, one is appended' with N_anc=6 implies the cache fills at latent frame 18, just inside L_pre=21. Stating this explicitly (and ablating N_anc and the stride) would clarify how much of the pretrained horizon the anchors span.
  4. [§Method, 'Efficiency'] The 16.5% latency overhead is reported for a single GPU (RTX A6000) without a throughput/latency table or breakdown between anchor attention and the 3D FFT/IFFT. A small table would strengthen the efficiency claim.
  5. [§Experiments, 'Evaluation'] The prompt-refinement citation appears mismatched: the text says prompts are refined by Qwen/Qwen2.5-7B-Instruct but cites Yang et al. 2025, which is the Qwen3 technical report.
  6. [§Error Accumulation in the Frequency Domain] The claim that Deep Forcing's spectral jitter 'manifests as abrupt temporal flickering' (§Error Accumulation) is supported only by Fig. 3; a quantitative flicker metric or a clearer side-by-side would help.
  7. [Fig. 4] Fig. 4 is hard to read at print size; consider fewer methods per row or larger crops, and indicate the timestamps of the sampled frames within the 120 s video.

Circularity Check

0 steps flagged

No significant circularity: SSA is an inference-time fusion method evaluated on external VBench metrics, not a result forced by its own inputs.

full rationale

FreqForcing’s load-bearing chain is observational characterization (spectral energy drift under Self-Forcing self-rollout; Fig. 3) → engineering intervention (dual-branch local/anchor attention fused by a fixed Gaussian low-pass in the frequency domain; Eqs. 7–12) → external benchmark comparison (VBench-Long on MovieGen prompts; Table 1). Nothing in that chain equates a claimed prediction to a fitted target by construction. Hyperparameters (σ=0.125, λ=0.6, N_anc=6, SSA only on the first two denoising steps) are chosen via ablation (Table 2) and then held fixed for the main comparison; that is ordinary method tuning, not a self-definitional or fitted-as-prediction loop. Citations (Self-Forcing, Rolling Forcing, FreeLong/FreeU, attention sink) supply prior techniques and baselines; none is a uniqueness theorem or ansatz by the present authors that forbids alternatives and thereby forces the result. The paper does not rename a known closed-form law as a derivation. Concerns about whether frozen early anchors differ from generic low-pass smoothing, or whether consistency metrics reward the injected similarity, are experimental-validity issues, not circularity. Score 0; steps empty.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claim rests on an empirical spectral diagnosis, a short list of hand-chosen inference hyperparameters, and standard AR diffusion / KV-cache assumptions. No new physical entities; the main inventions are methodological (SSA, anchor cache policy).

free parameters (5)
  • spectral mix strength λ = 0.6
    Controls how strongly low-frequency residual from anchor attention overwrites local attention; chosen by ablation on consistency vs repetitive rate.
  • Gaussian low-pass widths σ_xy, σ_t = 0.125
    Set the spatial/temporal frequency cutoff of the anchoring filter; swept jointly in Table 2 and fixed for main results.
  • anchor cache size N_anc = 6
    Number of frozen early frames used as spectral anchors; fixed by design/efficiency, not derived.
  • attention sink size S = 12
    Inherited and set to 12 following prior AR video practice; affects baseline drift and interacts with SSA.
  • SSA active denoising steps = first two of four steps
    SSA applied only at t=1000 and t=937 of the 4-step schedule; justified by coarse-to-fine intuition, not proven optimal.
axioms (4)
  • domain assumption Autoregressive self-rollout error accumulation is adequately characterized as DC/low-frequency spectral energy drift in attention outputs and pixels.
    Stated from STFT analysis on Self-Forcing 60 s rollouts (Fig. 3, Error Accumulation section); treated as the motivating fact for the whole method.
  • ad hoc to paper Early frames generated inside the pretrained horizon remain high-quality enough to serve as frozen low-frequency anchors for the full long horizon.
    Anchor cache fill-and-freeze policy (Eq. 8, N_anc=6) assumes early distribution match persists as a global reference.
  • domain assumption Sliding-window causal attention with KV cache is a sufficient approximation of full history for chunk-wise diffusion decoding.
    Standard Self-Forcing / AR video setup in Preliminaries (Eqs. 2–6); inherited without re-derivation.
  • domain assumption Low-frequency content of attention maps corresponds to global appearance/layout stability and high-frequency content to local motion/detail.
    Used to justify frequency-selective fusion (Spectral Anchoring, Eqs. 9–12); common in FreeU/FreeLong-style work but still an interpretive assumption here.
invented entities (2)
  • Spectral Self-Anchoring (SSA) operator no independent evidence
    purpose: Training-free fusion of local and anchor attention in the 3D frequency domain to suppress low-frequency drift.
    Core proposed mechanism; defined by dual branches plus Gaussian-gated residual in Eqs. 7–12. Methodological construct, not a physical entity.
  • Anchor KV cache (frozen early-frame bank) no independent evidence
    purpose: Supply stable keys/values from high-quality early frames once generation exceeds pretrained length.
    Policy artifact of the method (capacity 6, subsample every 3 latents, freeze after fill).

pith-pipeline@v1.2.0-grok45-kimik3 · 16768 in / 3478 out tokens · 63705 ms · 2026-07-30T11:28:38.692784+00:00 · methodology

0 comments
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 Jiatong Li, Leo Liang, Linghe Kong, Yulun Zhang.

Figure 1
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 l… view at source ↗
Figure 2
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 ge… view at source ↗
Figure 3
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 alle… view at source ↗
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 set for the i-th chunk is: bi = max (s + 1, if − w + s + 1), (3) Rlocal i = {n | bi ≤ n ≤ (i − 1)f} . (4) The final context with attention … view at source ↗
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 FreqForcing produces visually stable videos by preserving the spectral features. and to comprehensively understand this phenomenon, we provi… view at source ↗
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 al. 2023) ob￾serve that latent diffusion models synthesize different levels of visual content across denoising stages: the global scene lay… view at source ↗

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

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=

    Pith/arXiv 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=

    Pith/arXiv 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=

  9. [9]

    arXiv preprint arXiv:2509.25161 , year=

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

    Pith/arXiv arXiv

  10. [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=

    Pith/arXiv arXiv

  11. [11]

    NeurIPS , year=

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

  12. [12]

    arXiv preprint arXiv:2507.00162 , year=

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

    Pith/arXiv arXiv

  13. [13]

    ICLR , year=

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

  14. [14]

    CVPR , year=

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

  15. [15]

    ICML , year=

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

  16. [16]

    ICLR , year=

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

  17. [17]

    ICLR , year=

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

  18. [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=

    Pith/arXiv arXiv

  19. [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=

    arXiv

  20. [20]

    ICLR , year=

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

  21. [21]

    arXiv preprint arXiv:2605.06509 , year=

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

    Pith/arXiv arXiv

  22. [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=

    Pith/arXiv arXiv

  23. [23]

    ICML , year=

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

  24. [24]

    arXiv preprint arXiv:2505.13211 , year=

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

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2504.13074 , year=

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

    Pith/arXiv arXiv

  26. [26]

    NeurIPS , year=

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

  27. [27]

    CVPR , year=

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

  28. [28]

    arXiv preprint arXiv:2402.09470 , year=

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

    Pith/arXiv arXiv

  29. [29]

    ICLR , year=

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

  30. [30]

    ICLR , year=

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

  31. [31]

    ECCV , year=

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

  32. [32]

    ICLR , year=

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

  33. [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=

  34. [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=

  35. [35]

    arXiv preprint arXiv:2505.09388 , year=

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

    Pith/arXiv arXiv

  36. [36]

    arXiv preprint arXiv:2410.13720 , year=

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

    Pith/arXiv arXiv

  37. [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=

    Pith/arXiv arXiv

  38. [38]

    ICCV , year=

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

  39. [39]

    ICLR , year=

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

  40. [40]

    arXiv preprint arXiv:1404.1100 , year=

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

    Pith/arXiv arXiv

  41. [41]

    NeurIPS , year=

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

  42. [42]

    ICCV , year=

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

  43. [43]

    Neurocomputing , year=

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

  44. [44]

    arXiv preprint arXiv:2304.07193 , year=

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

    Pith/arXiv arXiv

  45. [45]

    ICLR , year=

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