The reviewed record of science sign in
Pith

arxiv: 2605.30174 · v1 · pith:4Y3LW4SQ · submitted 2026-05-28 · cs.CV

LiveSVG: Zero-Shot SVG Animation via Video Generation

Matan Levy , Ran Margolin , Bar Cavia , Dvir Samuel , Yael Pritch , Shmuel Peleg , Alex Rav Acha , Ariel Shamir
show 1 more author
Dani Lischinski
This is my paper

Reviewed by Pith2026-06-29 07:57 UTCgrok-4.3pith:4Y3LW4SQopen to challenge →

classification cs.CV
keywords SVG animationzero-shotvideo diffusiondifferentiable renderingBezier fittingvector graphicsmotion transfer
0
0 comments X

The pith

LiveSVG animates input SVGs by fitting their paths directly to a target video from a frozen image-to-video model.

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

The paper presents a zero-shot method for SVG animation that avoids LLM code synthesis and score distillation sampling. It first generates a previewable target video from the input SVG and a text motion prompt using an off-the-shelf image-to-video diffusion model. The original vector paths are then optimized to match this video through differentiable rendering. A dual-level motion model handles both rigid group movements and local non-rigid deformations while a recolorization step removes color-based matching ambiguities. This produces prompt-aligned, fully editable vector animations that outperform prior approaches on standard and new complex benchmarks.

Core claim

LiveSVG fits the input SVG directly to a target video generated by a frozen image-to-video model using a skeleton-free dual-level motion representation of homographies and Bezier offsets, along with sphere-packing recolorization, to achieve zero-shot prompt-aligned SVG animation.

What carries the argument

Dual-level motion representation combining per-group homographies for coarse articulation with per-path Bezier control-point offsets for local deformations, enabled by sphere-packing recolorization to resolve color-induced correspondence issues during pixel-wise fitting.

If this is right

  • Complex multi-object SVG scenes can be animated without category-specific priors such as skeletons.
  • The resulting animations remain fully editable vector graphics rather than raster outputs.
  • Direct reference-video fitting supplies cleaner gradients than SDS for non-rigid Bezier deformations.
  • The ChallengeSVG benchmark exposes where LLM synthesis and SDS methods fail on intricate motions.

Where Pith is reading between the lines

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

  • Replacing the image-to-video model with a stronger one would immediately improve animation fidelity without changing the fitting pipeline.
  • The same dual-level fitting could be applied to other vector representations such as fonts or UI elements.
  • If the per-frame optimization is accelerated, the method could support interactive SVG animation from live prompts.

Load-bearing premise

The video produced by the frozen image-to-video model supplies motion that the homography-plus-Bezier fitting can match accurately without skeletons or post-processing corrections.

What would settle it

Running the fitting stage on a generated video whose motion includes topology changes or deformations outside the span of per-group homographies and per-path Bezier offsets, then measuring whether the rendered output diverges from the video reference.

Figures

Figures reproduced from arXiv: 2605.30174 by Alex Rav Acha, Ariel Shamir, Bar Cavia, Dani Lischinski, Dvir Samuel, Matan Levy, Ran Margolin, Shmuel Peleg, Yael Pritch.

Figure 1
Figure 1. Figure 1: LiveSVG animates a static SVG from a motion prompt in a zero-shot setting. It first generates a previewable target video with an image-to-video model, then fits the original vector geometry directly to the target sequence. The leftmost image in each example is the input SVG, followed by three frames from the resulting editable SVG animation. Project page: https://levymsn.github.io/LiveSVG . We introduce Li… view at source ↗
Figure 2
Figure 2. Figure 2: Representative images and structural statistics for AniClipart and [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of LiveSVG. We first preprocess the input SVG with semantic grouping and sphere-packing recolorization, then generate candidate videos from the recolorized SVG and filter them to select the best target video. Finally, we fit each target frame back to the SVG by optimizing per-group homographies and per-path Bézier control-point deformations. Our constructed prompt explicitly asks for motions that … view at source ↗
Figure 5
Figure 5. Figure 5: Overall preference win rates on ChallengeSVG. Left: human user [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative breadth of LiveSVG across diverse prompts and SVG structures [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison for the prompt “A man sits on the floor”. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison on a full-split dancer prompt. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Seed diversity of LiveSVG. Variants A–D preserve the SVG structure while producing distinct plausible motions for the same input and prompt [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Automated A/B preference study results on ChallengeSVG with a Gemini-based judge. Left: per-criterion outcomes. Right: overall win rate against [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Automated A/B preference study results for the main AniClipart setting with a Gemini-based judge. Left: per-criterion outcomes. Right: overall win [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Automated A/B preference study results for [PITH_FULL_IMAGE:figures/full_fig_p013_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Human artifact ablation study on AniClipart. Workers choose which [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Amazon Mechanical Turk interface for the human A/B preference studies. Workers viewed the static SVG reference and two anonymized animations, [PITH_FULL_IMAGE:figures/full_fig_p015_14.png] view at source ↗
read the original abstract

We introduce LiveSVG, a zero-shot approach for generating Scalable Vector Graphics (SVG) animations using video diffusion models. Current SVG animation methods struggle with complex motions: LLM-based code synthesis fails to express fine, non-rigid B\'ezier deformations, while Score Distillation Sampling (SDS) provides noisy gradients and often requires category-specific priors like skeletons. In contrast, LiveSVG fits vector geometry directly to an explicitly generated target video. Given an input SVG image and a motion prompt, we generate a previewable target video using a frozen image-to-video model, then fit the original SVG to this video via differentiable rendering. Our fitting stage is skeleton-free, utilizing a dual-level motion representation that combines per-group homographies for coarse articulation with per-path B\'ezier control-point offsets for local deformations. To resolve color-induced correspondence ambiguities during pixel-wise fitting, we introduce a novel sphere-packing recolorization strategy. We also present ChallengeSVG, a benchmark of complex, multi-object scenes that exposes the limitations of prior work. Evaluations demonstrate that LiveSVG significantly outperforms existing methods on both AniClipart and ChallengeSVG, establishing direct reference-video fitting as a practical, robust route to prompt-aligned and fully editable vector animation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces LiveSVG, a zero-shot SVG animation method that conditions a frozen image-to-video model on an input SVG and motion prompt to generate a target video, then optimizes a dual-level motion representation (per-group homographies for coarse motion plus per-path Bézier control-point offsets for local deformations) via differentiable rendering to fit the SVG to the video; a sphere-packing recolorization strategy is used to resolve color ambiguities in pixel-wise correspondence. The work also introduces the ChallengeSVG benchmark for complex multi-object scenes and claims significant outperformance over prior LLM-synthesis and SDS-based methods on both AniClipart and ChallengeSVG.

Significance. If the fitting results hold under the stated representation, the approach offers a practical skeleton-free alternative for producing prompt-aligned, fully editable vector animations, bypassing noisy gradients from SDS and limited expressivity of LLM code synthesis; the new ChallengeSVG benchmark could also serve as a useful stress test for future methods handling non-rigid multi-object motion.

major comments (2)
  1. [fitting stage description and abstract] The central outperformance claim (abstract) that direct reference-video fitting is 'practical, robust' and resolves limitations of SDS/LLM methods is load-bearing on the unverified premise that motions produced by the frozen I2V model lie within the span of the dual-level representation (per-group homographies + per-path Bézier offsets). No analysis or failure-case quantification is provided on the fraction of ChallengeSVG examples where I2V outputs include non-rigid deformations, occlusions, or timing outside this span, which would prevent the optimizer from recovering correct parameters.
  2. [sphere-packing recolorization strategy] The sphere-packing recolorization strategy is introduced to resolve color-induced correspondence ambiguities during pixel-wise fitting, yet the description makes clear it operates only on color assignment; it does not address geometric correspondence after non-rigid motion, leaving open whether the optimizer can unambiguously recover the homography and Bézier parameters without skeletons or post-processing on ChallengeSVG cases.
minor comments (2)
  1. [abstract] Notation for Bézier is inconsistently escaped in the abstract; ensure consistent math formatting throughout.
  2. [abstract] The abstract states evaluations demonstrate 'significant' outperformance but does not preview any quantitative metrics, ablation details, or error bars; adding a brief summary of key numbers in the abstract would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We respond point-by-point to the major comments below.

read point-by-point responses

  1. Referee: [fitting stage description and abstract] The central outperformance claim (abstract) that direct reference-video fitting is 'practical, robust' and resolves limitations of SDS/LLM methods is load-bearing on the unverified premise that motions produced by the frozen I2V model lie within the span of the dual-level representation (per-group homographies + per-path Bézier offsets). No analysis or failure-case quantification is provided on the fraction of ChallengeSVG examples where I2V outputs include non-rigid deformations, occlusions, or timing outside this span, which would prevent the optimizer from recovering correct parameters.

    Authors: We agree that the manuscript would benefit from explicit discussion of the dual-level representation's coverage. The per-group homographies and per-path Bézier offsets are chosen to span the range of motions typically produced by the frozen I2V model on the evaluated prompts; quantitative results on ChallengeSVG show that the optimizer recovers parameters yielding lower error than baselines. In revision we will add a limitations subsection with failure-case examples (e.g., severe self-occlusion or timing mismatches) and a simple coverage statistic on the benchmark, making the scope of the claim transparent without altering the reported performance numbers. revision: partial

  2. Referee: [sphere-packing recolorization strategy] The sphere-packing recolorization strategy is introduced to resolve color-induced correspondence ambiguities during pixel-wise fitting, yet the description makes clear it operates only on color assignment; it does not address geometric correspondence after non-rigid motion, leaving open whether the optimizer can unambiguously recover the homography and Bézier parameters without skeletons or post-processing on ChallengeSVG cases.

    Authors: The sphere-packing step is strictly for color disambiguation to improve the reliability of the pixel-wise loss; geometric recovery occurs via direct optimization of the dual-level parameters against the rendered-versus-video image loss. Because the loss is image-based rather than correspondence-based, the optimizer can converge on plausible homography and offset values even when exact point-wise geometric matches are ambiguous. We will revise the method section to separate the color and geometric roles more clearly and will add qualitative optimization traces on ChallengeSVG examples to illustrate convergence behavior. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The derivation chain generates a target video from a frozen external image-to-video model conditioned on the input SVG and prompt, then optimizes dual-level parameters (group homographies + per-path Bézier offsets) via differentiable rendering to match that video. This fitting step produces the output animation but does not reduce any claimed result or prediction to a quantity defined by the method itself; the reference video is independent input. No self-citations, uniqueness theorems, or ansatzes from prior author work are invoked as load-bearing premises. The sphere-packing recolorization and ChallengeSVG benchmark are presented as novel contributions without circular reduction. The pipeline is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The central claim rests on the quality of the external video diffusion model and the effectiveness of the newly introduced fitting components; these are not derived from first principles but introduced as engineering solutions.

free parameters (2)
  • homography parameters per group
    Coarse articulation parameters optimized during fitting to the target video.
  • Bezier control-point offsets per path
    Local deformation parameters optimized to match video motion.
axioms (1)
  • domain assumption The target video generated by the frozen image-to-video model provides a reliable motion reference suitable for pixel-wise fitting.
    Invoked when the method states that fitting the SVG directly to the generated video resolves prior limitations.
invented entities (1)
  • sphere-packing recolorization strategy no independent evidence
    purpose: Resolve color-induced correspondence ambiguities during pixel-wise fitting
    New technique introduced to handle color matching issues in the fitting stage.

pith-pipeline@v0.9.1-grok · 5774 in / 1401 out tokens · 34379 ms · 2026-06-29T07:57:12.500121+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 10 canonical work pages · 4 internal anchors

  1. [1]

    Wen-Hui Du and Francis JM Schmitt

    Définition numérique des courbes et surfaces II.Automatismes12 (1967), 17–21. Wen-Hui Du and Francis JM Schmitt

    1967

  2. [2]

    Rinon Gal, Yael Vinker, Yuval Alaluf, Amit Bermano, Daniel Cohen-Or, Ariel Shamir, and Gal Chechik

    On the G1 continuity of piecewise Bézier surfaces: a review with new results.Computer-Aided Design22, 9 (1990), 556–573. Rinon Gal, Yael Vinker, Yuval Alaluf, Amit Bermano, Daniel Cohen-Or, Ariel Shamir, and Gal Chechik

    1990

  3. [3]

    In2025 IEEE International Conference on Image Processing, ICIP 2025 - Workshops, Anchorage, AK, USA, September 14-17,

    LINR Bridge: Vector Graphic Animation via Neural Implicits and Video Diffusion Priors. In2025 IEEE International Conference on Image Processing, ICIP 2025 - Workshops, Anchorage, AK, USA, September 14-17,

    2025

  4. [4]

    LTX-Video: Realtime Video Latent Diffusion

    LTX-Video: Realtime Video Latent Diffusion.CoRRabs/2501.00103 (2025). Mhand Hifi and Rym M’Hallah

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    A Literature Review on Circle and Sphere Packing Problems: Models and Methodologies.Adv. Oper. Res.(2009), 150624:1–150624:22. Ajay Jain, Amber Xie, and Pieter Abbeel

    2009

  6. [6]

    VectorFusion: Text-to-SVG by Abstract- ing Pixel-Based Diffusion Models. InCVPR. IEEE, 1911–1920. Anant Khandelwal

    1911

  7. [7]

    Graph.39, 6 (2020), 193:1–193:15

    Differen- tiable vector graphics rasterization for editing and learning.ACM Trans. Graph.39, 6 (2020), 193:1–193:15. Guotao Liang, Zhangcheng Wang, Chuang Wang, Juncheng Hu, Haitao Zhou, Jun- hua Liu, Jing Zhang, Dong Xu, and Qian Yu

    2020

  8. [8]

    VAnim: Rendering-Aware Sparse State Modeling for Structure-Preserving Vector Animation

    VAnim: Rendering-Aware Sparse State Modeling for Structure-Preserving Vector Animation.arXiv preprint arXiv:2605.01517(2026). Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  9. [9]

    Decomate: Lever- aging generative models for co-creative svg animation.arXiv preprint arXiv:2511.06297, 2025

    De- comate: Leveraging Generative Models for Co-Creative SVG Animation.CoRR abs/2511.06297 (2025). Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall

    work page arXiv 2025

  10. [10]

    Gemini: A Family of Highly Capable Multimodal Models

    Gemini: A Family of Highly Capable Multimodal Models.CoRR abs/2312.11805 (2023). Tiffany Tseng, Ruijia Cheng, and Jeffrey Nichols

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  11. [11]

    Keyframer: Empowering An- imation Design using Large Language Models.arXiv preprint arXiv:2402.06071 (2024). Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengy...

    work page arXiv 2024

  12. [12]

    Wan: Open and Advanced Large-Scale Video Generative Models.arXiv preprint arXiv:2503.20314(2025). Haomin Wang, Jinhui Yin, Qi Wei, Wenguang Zeng, Lixin Gu, Shenglong Ye, Zhangwei Gao, Yaohui Wang, Yanting Zhang, Yuanqi Li, Yanwen Guo, Wenhai Wang, Kai Chen, Yu Qiao, and Hongjie Zhang

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    World Wide Web Consortium

    InternSVG: Towards Unified SVG Tasks with Multimodal Large Language Models.ICLR(2026). World Wide Web Consortium

    2026

  14. [14]

    AniClipart: Clipart Anima- tion with Text-to-Video Priors.Int. J. Comput. Vis.133, 6 (2025), 3149–3165. Ximing Xing, Juncheng Hu, Guotao Liang, Jing Zhang, Dong Xu, and Qian Yu

    2025

  15. [15]

    Yiying Yang, Wei Cheng, Sijin Chen, Xianfang Zeng, Jiaxu Zhang, Liao Wang, Gang Yu, Xingjun Ma, and Yu-Gang Jiang

    OmniLottie: Generating Vector Animations via Parameterized Lottie Tokens.CoRRabs/2603.02138 (2026). Yiying Yang, Wei Cheng, Sijin Chen, Xianfang Zeng, Jiaxu Zhang, Liao Wang, Gang Yu, Xingjun Ma, and Yu-Gang Jiang

    work page arXiv 2026

  16. [16]

    Omnisvg: A unified scalable vector graphics genera- tion model.arXiv preprint arxiv:2504.06263, 2025

    OmniSVG: A Unified Scalable Vector Graphics Generation Model.CoRRabs/2504.06263 (2025). Jooyeol Yun and Jaegul Choo

    work page arXiv 2025

  17. [17]

    Artem Zholus, Carl Doersch, Yi Yang, Skanda Koppula, Viorica Patraucean, Xu Owen He, Ignacio Rocco, Mehdi S

    Vector Prism: Animating Vector Graphics by Stratifying Semantic Structure.CoRRabs/2512.14336 (2025). Artem Zholus, Carl Doersch, Yi Yang, Skanda Koppula, Viorica Patraucean, Xu Owen He, Ignacio Rocco, Mehdi S. M. Sajjadi, Sarath Chandar, and Ross Goroshin

    work page arXiv 2025

  18. [18]

    TAPNext: Tracking Any Point (TAP) as Next Token Prediction.CoRRabs/2504.05579 (2025). Fig

    work page arXiv 2025

  19. [19]

    On this subset,LiveSVGis the fastest GPU optimization method at 4.7 minutes per successful sample, while several baselines slow substantially on complex SVGs. This difference reflects the fact that methods such as Vector Prism and skeleton/template-heavy optimization baselines expose latency that depends on serialized SVG element complexity or structural ...

    2023

  20. [20]

    For optimization-based methods,boldmarks the best value and underlining marks the second best; ties share the same emphasis

    Quantitative comparison on the ChallengeSVG benchmark. For optimization-based methods,boldmarks the best value and underlining marks the second best; ties share the same emphasis. The VLM baseline generates animation programs directly; in our benchmark outputs, these programs are usually transform-centric and only rarely animate path ge- ometry explicitly...

    2023

  21. [21]

    Workers viewed the static SVG reference and two anonymized animations, then selected the animation they preferred overall

    Amazon Mechanical Turk interface for the human A/B preference studies. Workers viewed the static SVG reference and two anonymized animations, then selected the animation they preferred overall. answers on each duplicated pair after mapping A/B choices back to method identities. Workers with more than 50% inconsistent an- swers on these repeated comparison...

    2023

  22. [22]

    Main experimental configuration forLiveSVG. Component Setting Reference generator Veo 3.1 by default; Wan Video and LTX- Video for reference-generator variants Candidate generation 10–20 videos per input, sampled with dif- ferent random seeds Candidate selection Gemini two-stage scoring and comparative reranking using video clips and represen- tative keyf...

    2000

  23. [23]

    Component Model / toolkit Semantic grouping + cap- tioning Gemini 3.1 Pro [Team 2023] Reference video genera- tion Veo 3.1 [Google DeepMind 2024] (default), Wan Video [Wan et al

    External models and toolkits used inLiveSVG. Component Model / toolkit Semantic grouping + cap- tioning Gemini 3.1 Pro [Team 2023] Reference video genera- tion Veo 3.1 [Google DeepMind 2024] (default), Wan Video [Wan et al . 2025] and LTX- Video [HaCohen et al. 2025] (variants) Differentiable renderer DiffVG [Li et al. 2020] Tracking-based keyframe initia...

    2023

  24. [24]

    Exponential spatial regularization weights adjacent control-point offset differences by 𝑤 𝑖 𝑗 =exp[−(∥𝑥 0 𝑖 −𝑥 0 𝑗 ∥2/𝜎𝑠 )2] with 𝜎𝑠 = 0.01𝑊 , where𝑊 is the SVG canvas width

    and samples this map at raw control- point locations. Exponential spatial regularization weights adjacent control-point offset differences by 𝑤 𝑖 𝑗 =exp[−(∥𝑥 0 𝑖 −𝑥 0 𝑗 ∥2/𝜎𝑠 )2] with 𝜎𝑠 = 0.01𝑊 , where𝑊 is the SVG canvas width. For cubic Bézier joints (𝑐 − 𝑗 , 𝑎𝑗, 𝑐+ 𝑗 ), the 𝐺 1 loss is mean𝑘,𝑗 1−cos(𝑎 𝑗 −𝑐 − 𝑗 , 𝑐+ 𝑗 −𝑎 𝑗 ) , applied to joints that are...

    2026