Pith. sign in

REVIEW 5 major objections 5 minor 29 references

SketchAnimator: Animate Sketch via Motion Customization of Text-to-Video Diffusion Models

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read SketchAnimator turns a static vector sketch and a reference video into an animated sketch video, preserving the sketch's appearance while transferring the video's motion through a three-stage LoRA-and-SDS pipeline.

desk verdict A sensible three-stage pipeline combining motion LoRAs with SDS sketch optimization, but the evaluation and the untested additive LoRA composition keep it at 'promising, needs revision' rather than 'solidly demonstrated.' read the letter →

arxiv 2508.07149 v1 pith:CPRPCVIC submitted 2025-08-10 cs.CV

Ruolin Yang , Da Li , Honggang Zhang , Yi-Zhe Song This is my paper
classification cs.CV
keywords sketchanimationtext-to-videodiffusionmotioncustomizationLow-RankAdaptation(LoRA)ScoreDistillationSampling(SDS)Beziercurvesone-shotvideogenerationdifferentiablerasterization
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

The paper claims that a static sketch can be animated with the motion of any reference video in a one-shot setting, without key-point annotation or per-category training. It separates the task into learning the sketch's look, learning the video's motion, and then distilling both into the sketch's strokes. The key move is to fine-tune two Low-Rank Adaptation (LoRA) modules on a base video-diffusion model, one for the sketch and one for the video, and then add them into a single weight matrix that acts as a combined video prior. Score Distillation Sampling (SDS) then updates the Bezier-curve control points of the sketch, duplicated across frames, so the rendered frames match that prior. If the claim holds, amateurs can produce customized, creative sketch animations like "a jumping car" from a single reference video.

What carries the argument

The load-bearing object is the additive combination of two Low-Rank Adaptation (LoRA) updates inside a pre-trained video diffusion U-Net, Eq. (9): $W' = W_0 + \lambda_1 \Delta W_A + \lambda_2 \Delta W_M$. $\Delta W_A$ is trained on one static sketch image; $\Delta W_M$ is trained on one reference video. The summed weights act as the "video prior" whose denoising score is plugged into Score Distillation Sampling (SDS). SDS backpropagates through a differentiable Bezier rasterizer to adjust the control points of the sketch strokes, duplicated once per output frame. The claim is that this additive prior is coherent enough that SDS will drive the strokes into a video that preserves the sketch's

What would settle it

Generate directly from the combined model $W'$ (sketch LoRA + video LoRA) without SDS, using the sketch's semantic token and the motion caption, and check whether the sampled frames contain both the sketch's appearance and the reference motion. If the samples collapse to one side (photorealistic but not the sketch) or the other (static sketch), the additive prior that SDS is supposed to follow does not exist, and the distillation cannot be inheriting both signals.

Watch

Extended reading notes

Core claim

One-shot sketch animation, the paper argues, reduces to three decoupled stages on a frozen text-to-video model. Appearance LoRAs learn the sketch's identity from its semantic name; spatial and temporal LoRAs learn the reference video's look and dynamics; then the two LoRA updates are added into the base weights ($W' = W_0 + \lambda_1 \Delta W_A + \lambda_2 \Delta W_M$) to form a combined video prior. Score Distillation Sampling uses that prior to move the Bezier control points of the sketch, duplicated across frames and rendered by a differentiable rasterizer. On the MGIF benchmark the method reports the best appearance alignment (0.955), motion alignment (0.541), and temporal consistency (0

Load-bearing premise

The method assumes that adding a LoRA trained on one static sketch and a LoRA trained on one reference video produces a valid joint video prior, meaning neither update corrupts the other, so that score distillation can actually find strokes satisfying both.

Editorial extensions

If this is right

  • A user can animate any vector sketch with any reference video, without key-point annotations or per-category training.
  • The method can generate subject-motion combinations the pre-trained model has never seen, such as "a jumping car" or "a running clock".
  • Training requires only two 500-iteration LoRA stages plus SDS optimization, runnable on a single consumer GPU (RTX 3090).
  • Both components are necessary: removing the appearance LoRA degrades sketch identity; removing the motion LoRA makes the output static.
  • Compared methods that rely on key-point or pixel-level transfer deform the sketch or stay static; the staged decoupling is what lets the method balance appearance and motion.

Reading between the lines

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

  • The paper never verifies that the additive LoRA sum itself is a working video prior, only the full pipeline and single-LoRA ablations. A direct sampling test of $W'$ would settle whether the mechanism is genuinely additive or whether SDS is doing more of the work than the additive assumption implies.
  • SDS is known to push toward the mode of the score distribution; the reported motion alignment of 0.541 (vs. appearance 0.955) hints that the combined prior favors appearance, and the method may systematically under-translate fast or unusual motions.
  • The same two-LoRA-plus-SDS recipe should transfer beyond hand-drawn sketches: any parametric vector content (diagrams, logos, letters) could be animated by swapping the differentiable rasterizer and appearance LoRA.
  • A testable stress case is a cross-category pair, e.g., a tree sketch with a galloping-horse video. The paper's "creative motion" claim predicts a recognizable gallop; if the output instead deforms the tree or moves it randomly, the one-shot transfer is limited to near-category subjects.
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

5 major / 5 minor

Summary. The paper presents SketchAnimator, a three-stage pipeline for animating a user-provided vector sketch using motion from a single reference video. In Stage 1, appearance LoRAs are trained on the static sketch; in Stage 2, motion LoRAs are trained on the reference video while auxiliary A'-LoRAs are used in a spatial reconstruction loss; in Stage 3, both LoRAs are additively merged into the weights of a pretrained text-to-video diffusion model (Eq. 9), and score distillation sampling (SDS) updates duplicated Bezier curves to produce the output sketch video. The authors claim that the result retains the sketch's appearance while mirroring the reference video's motion, and report best scores on appearance alignment, motion alignment, and temporal consistency in Table I, with ablations removing each LoRA. The method builds on standard components (Modelscope, LoRA, diffvg, SDS) and is qualitatively compared against FOMM, Live Sketch, and several customized video generation methods.

Significance. If validated, SketchAnimator would provide a useful one-shot sketch animation tool that avoids manual keypoint annotation and extends motion customization to vector sketches. The decomposition into appearance and motion LoRAs followed by SDS-based Bezier optimization is a coherent and potentially practical design, and the paper addresses a real gap: previous sketch animation methods either require annotations, are domain-specific, or rely on text-only priors. The qualitative idea of combining one sketch-derived LoRA and one video-derived LoRA is appealing. However, the validation currently falls short of establishing the central claim: the motion metric is a text-based X-CLIP score rather than a comparison to the reference video, the appearance improvement over the ablation is very small, and the additive LoRA composition in Eq. (9) is asserted rather than directly tested. These issues are fixable with additional experiments, but they are load-bearing for the paper's main conclusion.

major comments (5)
  1. [§II-B, Eq. (9)] The central claim depends on W' = W0 + λ1ΔWA + λ2ΔWM, an additive combination of a LoRA trained on one static sketch image and a LoRA trained on one reference video. Since transformer attention layers are nonlinear, adding low-rank updates does not guarantee that the composed model behaves as a joint sketch-and-motion prior. The ablation in Table I removes one LoRA at a time, which only shows that each contributes, not that they compose correctly. Please provide a direct test of the composition, e.g., compare SDS optimization with W' against a jointly trained LoRA, or at least measure whether W' alone, without SDS, produces videos that exhibit both sketch identity and reference motion. Without such evidence, Eq. (9) remains an untested assumption.
  2. [§III-A, Motion Alignment metric] The Motion Alignment metric uses X-CLIP similarity between the generated video and the prompt "depicting reference video motion", not between the generated video and the reference video itself. Because the M-LoRA is trained on that same video with that prompt (Eqs. 6–7), and the SDS stage in Eq. (10) optimizes against the same text-conditioned model, a high score can partly reflect successful fitting to the training signal rather than correct motion transfer. Please add a motion metric computed against the reference video (e.g., optical flow, point trajectories, or a learned video-level similarity), report its correlation with human judgment, and clarify whether the X-CLIP score is meant as a semantic or a motion-fidelity measure.
  3. [Table I and §III-A] The quantitative evaluation has no error bars, no repeated runs, no significance tests, and the evaluation set is underdescribed: "For each video, we pick 5 sketches from CLIPasso, QuickDraw and SketchVOS" does not state how many videos, how many sketch-video pairs, or how the sketches were matched to videos. The appearance gain of Ours (0.955) over w/o A-LoRAs (0.947) is 0.008, which is likely within run-to-run noise; the temporal consistency of w/o A-LoRAs (0.995) is actually higher than Ours (0.988). Please report means and standard deviations over multiple random seeds and distinct inputs, and discuss the temporal-consistency trade-off.
  4. [§II-B, Motion Learning] The role of A'-LoRAs is unclear. They are trained with the spatial loss in Eq. (6) during motion learning, but the final weight composition in Eq. (9) uses ΔWA from appearance learning and ΔWM from motion learning. Are A'-LoRAs discarded after motion learning? If they are used to prevent appearance leakage into M-LoRAs, that should be stated explicitly and included in the ablation; if they are not used, it is unclear why they are trained. Please clarify the parameter flow and whether Eq. (7) updates A'-LoRAs, M-LoRAs, or both.
  5. [Eqs. (4) and (10)] The SDS gradient as written, w(t)(ϵW(zt, τθ(y)), t) ∂x/∂ϕ, omits the sampled noise ϵ and the expected residual term; standard SDS is w(t)(ϵθ(zt, y, t) − ϵ) ∂x/∂ϕ. This is either a typographical error or an incorrect optimization objective. Since SDS is the core of the video prior distillation stage, please correct the formula and ensure the implementation matches the corrected expression.
minor comments (5)
  1. [§III-A, Evaluation metrics] “the generated video frames and the inference prompt” should read “the generated video frames with the inference prompt” or similar. Also, Table I has inconsistent spacing in column headers.
  2. [§II-B, Eq. (7)] The objective Lmotion = Lspatial + Ltemporal is stated without specifying which parameters (A'-LoRAs, M-LoRAs, or both) receive gradients from each term. Please define the trainable sets explicitly.
  3. [References] Reference [22] is cited as the MGIF dataset, but the listed title “Animating arbitrary objects via deep motion transfer” may not correspond to a dataset named MGIF; please verify the dataset source and citation.
  4. [Fig. 1] The prompt templates “A [V] is moving.” and “A horse is moving.” are shown in the figure but not fully explained in the text. Clarify how the motion prompt template is constructed for a given reference video and whether it is used as the conditioning prompt in Eq. (10).
  5. [§III-B, Fig. 2] The text refers to the “8th row” of Fig. 2, but rows are not numbered in the figure. Consider labeling rows or referring to subfigures explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the three-stage pipeline is independently evaluated against external baselines and its central claims are testable, not definitionally forced by the inputs.

full rationale

SketchAnimator divides the task into appearance learning (A-LoRA on the input sketch), motion learning (M-LoRA and A'-LoRA on the reference video), and video-prior distillation (SDS update of Bezier strokes under the combined LoRA weights). The appearance and motion properties are learned from the two inputs, but the output sketch video is not equal to a fitted value: the SDS optimization in Eq. (10) must still solve a nontrivial stroke-parameter search, and the ablations show that removing either LoRA degrades the corresponding metric. The Motion Alignment metric is an externally defined X-CLIP score against a text prompt, not the reference video itself or a training objective, so it does not by construction reward the fitted motion. The additive weight combination in Eq. (9) is a design assumption that could be wrong, but it is not circular; it is empirically tested through comparisons and ablations. The only self-citation, [24], is used as a dataset source for SketchVOS sketches and is not load-bearing on the method's derivation. No step reduces to its own input by definition, and no self-citation is used to forbid alternatives or to justify the central premise.

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

The free parameters are hand-set hyperparameters; the load-bearing assumption is additive LoRA combination. No unusual invented entities; the '[V]' token is a standard textual-inversion placeholder.

free parameters (5)
  • lambda1 (appearance LoRA weight) = 0.5
    Hand-set in Eq. (9); the output depends on the balance between appearance and motion LoRAs. No sensitivity analysis.
  • lambda2 (motion LoRA weight) = 1.0
    Hand-set in Eq. (9); no tuning analysis provided.
  • Rasterizer learning rate = 2.0e-3
    Optimization schedule for SDS updates of Bezier control points; chosen without reported sensitivity analysis.
  • LoRA training iterations = 500 per stage
    Chosen following [20]; no ablation on iteration counts.
  • Motion prompt template = e.g., 'A [V] is moving.'
    The text condition for motion distillation is authored per example; not automated or analyzed.
assumptions (5)
  • domain assumption The SDS gradient defined via a pre-trained video diffusion model provides a valid descent direction for optimizing Bezier stroke parameters (Eq. 10).
    Inherited from DreamFusion; the paper does not validate this for video diffusion plus vector sketches beyond its own results.
  • ad hoc to paper LoRA weights trained independently on a single sketch image and on a single reference video combine additively as W0 + lambda1*DeltaWA + lambda2*DeltaWM without catastrophic interference (Eq. 9).
    This is the paper's core mechanism and is asserted, not derived. No interference or cross-domain transfer analysis.
  • domain assumption The X-CLIP score between generated frames and a prompt describing the reference video measures motion alignment for sketch videos.
    The metric is adopted without validation for abstract sketch content or fine-grained motion similarity (Table I).
  • domain assumption The VAE latent and 3D U-Net of Modelscope adequately represent and denoise sketch video content after LoRA adaptation.
    Standard transfer assumption underlying Section II.B; not directly tested.
  • standard math The differentiable rasterizer (diffvg) gradients are usable for SDS optimization at stroke level.
    Established by prior work [12]-[14], [19]; assumed here without modification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SketchAnimator: Animate Sketch via Motion Customization of Text-to-Video Diffusion Models." pith.science (2026). https://pith.science/paper/CPRPCVIC

@misc{pith2026250807149,
  author       = {Pith},
  title        = {Pith review of: SketchAnimator: Animate Sketch via Motion Customization of Text-to-Video Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CPRPCVIC}},
  note         = {Machine review of arXiv:2508.07149}
}
read the original abstract

Sketching is a uniquely human tool for expressing ideas and creativity. The animation of sketches infuses life into these static drawings, opening a new dimension for designers. Animating sketches is a time-consuming process that demands professional skills and extensive experience, often proving daunting for amateurs. In this paper, we propose a novel sketch animation model SketchAnimator, which enables adding creative motion to a given sketch, like "a jumping car''. Namely, given an input sketch and a reference video, we divide the sketch animation into three stages: Appearance Learning, Motion Learning and Video Prior Distillation. In stages 1 and 2, we utilize LoRA to integrate sketch appearance information and motion dynamics from the reference video into the pre-trained T2V model. In the third stage, we utilize Score Distillation Sampling (SDS) to update the parameters of the Bezier curves in each sketch frame according to the acquired motion information. Consequently, our model produces a sketch video that not only retains the original appearance of the sketch but also mirrors the dynamic movements of the reference video. We compare our method with alternative approaches and demonstrate that it generates the desired sketch video under the challenge of one-shot motion customization.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 21 canonical work pages

  1. [1]

    Motiondirector: Motion customization of text-to-video diffusion models,

    R. Zhao, Y . Gu, J. Z. Wu, D. J. Zhang, J. Liu, W. Wu, J. Keppo, and M. Z. Shou, “Motiondirector: Motion customization of text-to-video diffusion models,” arXiv preprint arXiv:2310.08465 , 2023

    arXiv 2023

  2. [2]

    Dreamvideo: Composing your dream videos with customized subject and motion,

    Y . Wei, S. Zhang, Z. Qing, H. Yuan, Z. Liu, Y . Liu, Y . Zhang, J. Zhou, and H. Shan, “Dreamvideo: Composing your dream videos with customized subject and motion,” in in Proc. IEEE/CVF Conf. Comput. Vis. and Pattern Recognit. , 2024

    work page 2024

  3. [3]

    Customize-a-video: One-shot motion customization of text-to-video diffusion models,

    Y . Ren, Y . Zhou, J. Yang, J. Shi, D. Liu, F. Liu, M. Kwon, and A. Shrivastava, “Customize-a-video: One-shot motion customization of text-to-video diffusion models,” arXiv preprint arXiv:2402.14780 , 2024

    arXiv 2024

  4. [4]

    Live sketch: Video-driven dynamic deformation of static drawings,

    Q. Su, X. Bai, H. Fu, C.-L. Tai, and J. Wang, “Live sketch: Video-driven dynamic deformation of static drawings,” in in Proc. 2018 CHI Conf. Human Factors Comput. Syst. , pp. 1–12, 2018

    work page 2018

  5. [5]

    First order motion model for image animation,

    A. Siarohin, S. Lathuili `ere, S. Tulyakov, E. Ricci, and N. Sebe, “First order motion model for image animation,” in in Proc. Adv. Neural Inf. Proces. Syst., vol. 32, 2019

    work page 2019

  6. [6]

    Motion representations for articulated animation,

    A. Siarohin, O. J. Woodford, J. Ren, M. Chai, and S. Tulyakov, “Motion representations for articulated animation,” in in Proc. IEEE/CVF Conf. Comput. Vis. and Pattern Recognit. , pp. 13653–13662, 2021

    work page 2021

  7. [7]

    Thin-plate spline motion model for image animation,

    J. Zhao and H. Zhang, “Thin-plate spline motion model for image animation,” in in Proc. IEEE/CVF Conf. Comput. Vis. and Pattern Recognit., pp. 3657–3666, 2022

    work page 2022

  8. [8]

    Depth-aware generative adversarial network for talking head video generation,

    F.-T. Hong, L. Zhang, L. Shen, and D. Xu, “Depth-aware generative adversarial network for talking head video generation,” in in Proc. IEEE/CVF Conf. Comput. Vis. and Pattern Recognit. , pp. 3397–3406, 2022

    work page 2022

Show all 29 references
  1. [9]

    Motion and appearance adaptation for cross-domain motion transfer,

    B. Xu, B. Wang, J. Deng, J. Tao, T. Ge, Y . Jiang, W. Li, and L. Duan, “Motion and appearance adaptation for cross-domain motion transfer,” in in Proc. Eur . Conf. Comput. Vis. , pp. 529–545, Springer, 2022

    2022

  2. [10]

    Breathing life into sketches using text-to-video priors,

    R. Gal, Y . Vinker, Y . Alaluf, A. Bermano, D. Cohen-Or, A. Shamir, and G. Chechik, “Breathing life into sketches using text-to-video priors,” in in Proc. IEEE/CVF Conf. Comput. Vis. and Pattern Recognit. , pp. 4325– 4336, 2024

    2024

  3. [11]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” arXiv preprint arXiv:2106.09685 , 2021

    2021 arXiv

  4. [12]

    Clipasso: Semantically-aware object sketching,

    Y . Vinker, E. Pajouheshgar, J. Y . Bo, R. C. Bachmann, A. H. Bermano, D. Cohen-Or, A. Zamir, and A. Shamir, “Clipasso: Semantically-aware object sketching,” ACM Transact. Graph. (TOG) , vol. 41, no. 4, pp. 1– 11, 2022

    2022

  5. [13]

    Clipascene: Scene sketching with different types and levels of abstraction,

    Y . Vinker, Y . Alaluf, D. Cohen-Or, and A. Shamir, “Clipascene: Scene sketching with different types and levels of abstraction,” in in Proc. IEEE/CVF Int. Conf. Comput. Vis. , pp. 4146–4156, 2023

    2023

  6. [14]

    Diffsketcher: Text guided vector sketch synthesis through latent diffusion models,

    X. Xing, C. Wang, H. Zhou, J. Zhang, Q. Yu, and D. Xu, “Diffsketcher: Text guided vector sketch synthesis through latent diffusion models,” in Proc. Adv. Neural Inf. Proces. Syst. , vol. 36, 2024

    2024

  7. [15]

    Dreamfusion: Text- to-3d using 2d diffusion,

    B. Poole, A. Jain, J. T. Barron, and B. Mildenhall, “Dreamfusion: Text- to-3d using 2d diffusion,” arXiv preprint arXiv:2209.14988 , 2022

    2022 arXiv

  8. [16]

    Mod- elscope text-to-video technical report,

    J. Wang, H. Yuan, D. Chen, Y . Zhang, X. Wang, and S. Zhang, “Mod- elscope text-to-video technical report,” arXiv preprint arXiv:2308.06571, 2023

    2023 arXiv

  9. [17]

    Videocrafter1: Open diffusion models for high-quality video generation,

    H. Chen, M. Xia, Y . He, Y . Zhang, X. Cun, S. Yang, J. Xing, Y . Liu, Q. Chen, X. Wang, et al. , “Videocrafter1: Open diffusion models for high-quality video generation,” arXiv preprint arXiv:2310.19512 , 2023

    2023 arXiv

  10. [18]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” in Proc. Adv. Neural Inf. Proces. Syst. , vol. 34, pp. 8780– 8794, 2021

    2021

  11. [19]

    Differentiable vector graphics rasterization for editing and learning,

    T.-M. Li, M. Luk ´aˇc, M. Gharbi, and J. Ragan-Kelley, “Differentiable vector graphics rasterization for editing and learning,” ACM Transact. Graph. (TOG), vol. 39, no. 6, pp. 1–15, 2020

    2020

  12. [20]

    Dreambooth: Fine tuning text-to-image diffusion models for subject- driven generation,

    N. Ruiz, Y . Li, V . Jampani, Y . Pritch, M. Rubinstein, and K. Aberman, “Dreambooth: Fine tuning text-to-image diffusion models for subject- driven generation,” in in Proc. Adv. Neural Inf. Proces. Syst. , pp. 22500– 22510, 2023

    2023

  13. [21]

    An image is worth one word: Personalizing text-to- image generation using textual inversion,

    R. Gal, Y . Alaluf, Y . Atzmon, O. Patashnik, A. H. Bermano, G. Chechik, and D. Cohen-or, “An image is worth one word: Personalizing text-to- image generation using textual inversion,” in in Proc. Eleventh Int. Conf. Learning Represent., 2023

    2023

  14. [22]

    Animat- ing arbitrary objects via deep motion transfer,

    A. Siarohin, S. Lathuili `ere, S. Tulyakov, E. Ricci, and N. Sebe, “Animat- ing arbitrary objects via deep motion transfer,” in in Proc. IEEE/CVF Conf. Comput. Vis. and Pattern Recognit. , pp. 2377–2386, 2019

    2019

  15. [23]

    Quickdraw: improving drawing experience for geometric diagrams,

    S. Cheema, S. Gulwani, and J. LaViola, “Quickdraw: improving drawing experience for geometric diagrams,” in in Proc. SIGCHI Conf. Human Factors Comput. Syst. , pp. 1037–1064, 2012

    2012

  16. [24]

    Sketch-based video object segmentation: Benchmark and analysis,

    R. Yang, D. Li, C. Hu, T. Hospedales, H. Zhang, and Y .-Z. Song, “Sketch-based video object segmentation: Benchmark and analysis,” in in Proc. 34th British Machine Vis. Conf. , 2023

    2023

  17. [25]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al., “Learning transferable visual models from natural language supervision,” in in Proc. Int. Conf. machine learning , pp. 8748–8763, 2021

    2021

  18. [26]

    Expanding language-image pretrained models for general video recognition,

    B. Ni, H. Peng, M. Chen, S. Zhang, G. Meng, J. Fu, S. Xiang, and H. Ling, “Expanding language-image pretrained models for general video recognition,” in in Proc. Eur . Conf. Comput. Vis., pp. 1–18, 2022

    2022

  19. [27]

    Runway. gen-2: Text driven video generation

    “Runway. gen-2: Text driven video generation..” https://research. runwayml.com/gen2, 2023

    2023

  20. [28]

    Dynamicrafter: Animating open-domain images with video diffusion priors,

    J. Xing, M. Xia, Y . Zhang, H. Chen, X. Wang, T.-T. Wong, and Y . Shan, “Dynamicrafter: Animating open-domain images with video diffusion priors,” arXiv preprint arXiv:2310.12190 , 2023

    2023 arXiv

  21. [29]

    Stable video diffusion: Scaling latent video diffusion models to large datasets,

    A. Blattmann, T. Dockhorn, S. Kulal, D. Mendelevitch, M. Kilian, D. Lorenz, Y . Levi, Z. English, V . V oleti, A. Letts,et al. , “Stable video diffusion: Scaling latent video diffusion models to large datasets,” arXiv preprint arXiv:2311.15127, 2023

    2023 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.