pith. machine review for the scientific record. sign in

arxiv: 2605.01517 · v1 · submitted 2026-05-02 · 💻 cs.CV

Recognition: unknown

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

Chuang Wang, Dong Xu, Guotao Liang, Haitao Zhou, Jing Zhang, Juncheng Hu, Junhua Liu, Qian Yu, Zhangcheng Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-09 14:02 UTC · model grok-4.3

classification 💻 cs.CV
keywords SVG animationtext-to-vectorsparse state updatesstructure preservationreinforcement learningvector graphicsmotion planning
0
0 comments X

The pith

VAnim generates text-to-SVG animations through sparse updates to a fixed vector structure instead of full sequence regeneration.

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

The paper presents a framework that reframes SVG animation creation from text as a process of making targeted changes to an existing vector graphics file rather than building new code from scratch each time. This approach keeps the underlying document structure intact by design, which matters because professional designs rely on editable, resolution-independent files that maintain consistent topology even during motion. The method combines an initial planning step that identifies specific visual elements with a reinforcement learning loop that uses visual feedback to refine the code changes. Experiments show the resulting animations match input descriptions more closely and remain structurally valid compared with prior approaches that either scramble the file or limit motion to rigid transforms. A new dataset of 134k examples supports testing these capabilities across open-domain instructions.

Core claim

Animation is reconceptualized as Sparse State Updates on a persistent SVG DOM tree rather than sequence generation. This yields sequence compression exceeding 9.8 times while preserving the full DOM structure and all non-participating elements by construction. An Identification-First Motion Planning step grounds text instructions to explicit visual entities, and Rendering-Aware Reinforcement Learning via Group Relative Policy Optimization aligns the discrete code updates with high-fidelity visual signals from a video perception model, overcoming the non-differentiable rendering barrier.

What carries the argument

Sparse State Updates (SSU) on a persistent SVG DOM tree, which performs targeted code modifications while leaving the overall structure and untouched elements unchanged.

If this is right

  • Generated animations remain fully editable as standard SVG files with preserved topology and non-rigid motion capability.
  • Generation becomes more efficient through sequence length reduction while still supporting complex open-domain instructions.
  • Semantic alignment improves because updates are grounded to specific visual entities before code changes occur.
  • Professional workflows benefit from resolution-independent output that does not require post-processing to restore structure.
  • The same persistent-structure approach can handle both rigid and non-rigid deformations without separate pipelines.

Where Pith is reading between the lines

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

  • Design tools could add direct natural-language editing of existing SVG files by reusing the same sparse-update loop.
  • The compression benefit may extend to other vector formats if the DOM persistence pattern is adapted.
  • Interactive preview loops become feasible because each update can be rendered and checked without regenerating the entire file.
  • Longer animation sequences could be composed by chaining multiple sparse-update stages on the same base structure.

Load-bearing premise

The hybrid visual reward from a video perception model will correctly steer discrete SVG code changes toward faithful motion without misalignment or entity errors in the initial planning step.

What would settle it

A set of test cases with non-rigid deformations where the output SVG either loses topological consistency, alters non-participating elements, or visually diverges from the motion described in the input text.

Figures

Figures reproduced from arXiv: 2605.01517 by Chuang Wang, Dong Xu, Guotao Liang, Haitao Zhou, Jing Zhang, Juncheng Hu, Junhua Liu, Qian Yu, Zhangcheng Wang.

Figure 1
Figure 1. Figure 1: VAnim Generation Gallery. We present diverse vector animations generated from prompts. The results demonstrate VAnim’s capability to execute complex semantic instructions, ranging from non-rigid deformation and sequential logic to multi-object interaction, all while maintaining strict topological consistency. Please refer to the supplementary webpage for full video demonstrations. Abstract Scalable Vector … view at source ↗
Figure 2
Figure 2. Figure 2: Construction Pipeline of SVGAnim-134k. The process operates in three stages: (1) Structural Standardization: Raw Lottie files are rendered into ID-anchored SVG DOM trees and normalized to a unified coordinate space. (2) Sparse Encoding: The animation sequence is compressed into Sparse State Updates, represented by serialized differential tokens (visualized in purple). (3) Dual-Stream Annotation: An MLLM ge… view at source ↗
Figure 3
Figure 3. Figure 3: Token Efficiency Analysis. A comparison of cumulative token consumption over a 24-frame sequence. Naive frame-by￾frame generation (red dashed line) suffers from context explosion, reaching ∼86k tokens. In contrast, VAnim’s Sparse State Update mechanism (blue solid line) maintains a compact representation (∼9.2k tokens), achieving a 9.86× compression ratio. This ef￾ficiency is critical for enabling long-hor… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the VAnim Framework. The generation process is decomposed into two stages: (1) Identification-First Motion Planning, where the Multimodal LLM explicitly grounds visual entities to persistent SVG IDs (e.g., binding the object inside the orange box to id=06 through reasoning) and reasons about causal motion logic; (2) Sparse State Update, where the model predicts attribute differentials (Dt) only… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Comparison on SVGAnim-Test. We compare VAnim against state-of-the-art baselines on four challenging prompts requiring precise control and non-rigid deformation. Baselines suffer from severe Identity Drift and Topological Collapse. Specifically, GPT-5.2 changes the toucan’s wing color to red; Gemini distorts the bottle shape; while LiveSketch produces broken lines. Our VAnim generates smooth, se… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative Ablation. Row 1 (w/o Structure-Bound CoT) suffers from grounding failures, manipulating wrong en￾tities (e.g., rotating the whole cabinet instead of the door). Row 2 (w/o RL) exhibits motion laziness, producing conservative up￾dates that fail to fully satisfy the “close” or “open” instructions. Row 3 (Full) achieves accurate semantic alignment with precise deformation. causes a semantic drop (0… view at source ↗
read the original abstract

Scalable Vector Graphics (SVG) animation generation is pivotal for professional design due to their structural editability and resolution independence. However, this task remains challenging as it requires bridging discrete code representations with continuous visual dynamics. Existing optimization-based methods often destroy topological consistency, while general-purpose LLMs rely on rigid CSS/SMIL transformations, failing to model geometry-level non-rigid deformations. To address these limitations, we present VAnim, the first LLM-based framework for open-domain text-to-SVG animation. We reconceptualize animation not as sequence generation, but as Sparse State Updates (SSU) on a persistent SVG DOM tree. This paradigm compresses sequence length by over 9.8x while preserving the SVG DOM structure and non-participating elements by construction. To enable precise control, we propose an Identification-First Motion Planning mechanism that grounds textual instructions in explicit visual entities. Furthermore, to overcome the non-differentiable nature of SVG rendering, we employ Rendering-Aware Reinforcement Learning via Group Relative Policy Optimization (GRPO). By leveraging a hybrid reward from a state-of-the-art video perception encoder, we align discrete code updates with high-fidelity visual feedback. We also introduce SVGAnim-134k, the first benchmark for vector animation. Extensive experiments demonstrate that VAnim significantly outperforms state-of-the-art baselines in semantic alignment and structural validity, with additional appendix metrics further validating motion quality and identity preservation.

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 / 1 minor

Summary. The manuscript presents VAnim, an LLM-based framework for open-domain text-to-SVG animation. It reconceptualizes animation generation as Sparse State Updates (SSU) on a persistent SVG DOM tree, which compresses sequence length by over 9.8x while preserving structure and non-participating elements by construction. The approach includes Identification-First Motion Planning to ground textual instructions in visual entities and Rendering-Aware Reinforcement Learning via Group Relative Policy Optimization (GRPO) that uses a hybrid reward from a state-of-the-art video perception encoder to align discrete code updates with visual feedback. The paper introduces the SVGAnim-134k benchmark and reports that VAnim significantly outperforms state-of-the-art baselines in semantic alignment and structural validity.

Significance. If the results hold, this work could meaningfully advance controllable vector animation by preserving the editability and resolution independence of SVG while enabling non-rigid deformations beyond rigid CSS/SMIL transforms. The SSU paradigm and the new SVGAnim-134k benchmark are clear strengths that address gaps in structure preservation and evaluation. The rendering-aware RL component is a creative attempt to handle non-differentiability, though its impact depends on the fidelity of the visual reward signal.

major comments (2)
  1. [Rendering-Aware Reinforcement Learning via GRPO] Rendering-Aware Reinforcement Learning via GRPO section: The central performance claims rest on the hybrid reward from the raster-trained video perception encoder successfully guiding discrete SVG code updates toward geometric fidelity. Given the raster-to-vector domain gap, the manuscript should supply concrete evidence (e.g., correlation between reward values and geometric metrics such as path topology or control-point displacement) that the policy optimizes for true structural validity rather than rendering artifacts.
  2. [Experiments] Experiments section: The asserted outperformance in semantic alignment and structural validity requires explicit reporting of quantitative metrics, chosen baselines, error bars, and statistical tests. Without these details, it is difficult to assess whether the gains are robust or attributable to the proposed components.
minor comments (1)
  1. [Abstract] Abstract: Including one or two key quantitative results (e.g., specific improvement margins on the benchmark) would better substantiate the performance claims for readers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments and for recognizing the potential of the SSU paradigm and the new benchmark. We address the two major comments below and plan to revise the manuscript to incorporate additional evidence and clarifications.

read point-by-point responses

  1. Referee: Rendering-Aware Reinforcement Learning via GRPO section: The central performance claims rest on the hybrid reward from the raster-trained video perception encoder successfully guiding discrete SVG code updates toward geometric fidelity. Given the raster-to-vector domain gap, the manuscript should supply concrete evidence (e.g., correlation between reward values and geometric metrics such as path topology or control-point displacement) that the policy optimizes for true structural validity rather than rendering artifacts.

    Authors: We acknowledge the referee's valid concern regarding the raster-to-vector domain gap and the need for evidence that the reward guides geometric fidelity. The manuscript describes the hybrid reward in the Rendering-Aware RL section but does not provide explicit correlation analysis with geometric metrics. To address this, we will include in the revision a quantitative study reporting correlations between the reward signals and geometric metrics like path topology and control-point displacement on the SVGAnim-134k dataset. This will demonstrate that the policy optimizes for structural aspects beyond rendering artifacts. revision: yes

  2. Referee: Experiments section: The asserted outperformance in semantic alignment and structural validity requires explicit reporting of quantitative metrics, chosen baselines, error bars, and statistical tests. Without these details, it is difficult to assess whether the gains are robust or attributable to the proposed components.

    Authors: We agree that explicit details are necessary for assessing the robustness of our results. While the Experiments section reports outperformance with quantitative metrics on semantic alignment and structural validity using the introduced benchmark and several baselines, we will revise to explicitly detail the metrics (e.g., specific similarity scores and validity rates), list all chosen baselines, include error bars from repeated experiments, and add statistical tests such as significance levels. These enhancements will be made in the main text and tables. revision: yes

Circularity Check

0 steps flagged

No circularity: novel framework components and experimental claims remain independent of inputs

full rationale

The paper introduces VAnim as a new reconceptualization of SVG animation as Sparse State Updates (SSU) on a persistent DOM tree, plus Identification-First Motion Planning and Rendering-Aware RL via GRPO with a hybrid video-encoder reward. The abstract explicitly states the 9.8x compression and DOM preservation occur 'by construction,' which is definitional rather than a derived prediction. No equations, self-citations as load-bearing uniqueness theorems, fitted parameters renamed as predictions, or ansatzes smuggled via prior work appear in the provided text. Performance claims rest on a new benchmark (SVGAnim-134k) and comparisons to external baselines, keeping the derivation self-contained against external validation.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 3 invented entities

The central claims rest on the effectiveness of the new SSU paradigm, the grounding mechanism, and the perceptual reward model; no numerical free parameters are mentioned, but two domain assumptions about non-differentiability and reward reliability are required.

axioms (2)
  • domain assumption SVG rendering is non-differentiable, so direct gradient-based optimization is impossible and RL must be used instead.
    Explicitly stated as the reason for employing Rendering-Aware Reinforcement Learning via GRPO.
  • domain assumption A state-of-the-art video perception encoder supplies a hybrid reward that aligns discrete code changes with human-perceived animation quality.
    Used to train the policy without direct visual supervision on SVG output.
invented entities (3)
  • Sparse State Updates (SSU) no independent evidence
    purpose: Model animation as targeted updates on a persistent SVG DOM tree rather than full sequence generation.
    Core reconceptualization that compresses sequence length and preserves structure by construction.
  • Identification-First Motion Planning no independent evidence
    purpose: Ground textual instructions in explicit visual entities before planning motion.
    Proposed mechanism for precise control over which parts of the SVG change.
  • Rendering-Aware Reinforcement Learning via GRPO no independent evidence
    purpose: Train the model using visual feedback from a video encoder despite non-differentiable rendering.
    Training procedure that replaces direct supervision with perceptual rewards.

pith-pipeline@v0.9.0 · 5578 in / 1755 out tokens · 31602 ms · 2026-05-09T14:02:12.005988+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. BoostAPR: Boosting Automated Program Repair via Execution-Grounded Reinforcement Learning with Dual Reward Models

    cs.AI 2026-05 unverdicted novelty 7.0

    BoostAPR improves automated program repair by using execution-grounded RL with a sequence-level assessor and line-level credit allocator, reaching 40.7% on SWE-bench Verified and strong cross-language results.

  2. BoostAPR: Boosting Automated Program Repair via Execution-Grounded Reinforcement Learning with Dual Reward Models

    cs.AI 2026-05 unverdicted novelty 6.0

    BoostAPR uses supervised fine-tuning on verified fixes, dual sequence- and line-level reward models from execution feedback, and PPO to reach 40.7% on SWE-bench Verified with strong cross-language results.

  3. BoostAPR: Boosting Automated Program Repair via Execution-Grounded Reinforcement Learning with Dual Reward Models

    cs.AI 2026-05 unverdicted novelty 6.0

    BoostAPR boosts automated program repair by training a sequence-level assessor and line-level credit allocator from execution outcomes, then applying them in PPO to reach 40.7% on SWE-bench Verified.

Reference graph

Works this paper leans on

287 extracted references · 50 canonical work pages · cited by 1 Pith paper · 29 internal anchors

  1. [1]

    World Wide Web Consortium , title=

  2. [2]

    Proceedings of the eleventh international conference on 3D web technology , pages=

    Integrating web 2D and 3D technologies for architectural visualization: applications of SVG and X3D/VRML in environmental behavior simulation , author=. Proceedings of the eleventh international conference on 3D web technology , pages=

  3. [3]

    2005 , publisher=

    Visualizing Information Using SVG and X3D: XML-based Technologies for the XML-based Web , author=. 2005 , publisher=

    2005

  4. [4]

    科学技术与工程 , volume=

    面向汉字矢量图形特征的字向量表征方法 , author=. 科学技术与工程 , volume=

  5. [5]

    Auto-Encoding Variational Bayes

    Auto-encoding variational bayes , author=. arXiv preprint arXiv:1312.6114 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Communications of the ACM , volume=

    Generative adversarial networks , author=. Communications of the ACM , volume=. 2020 , publisher=

    2020

  7. [7]

    Proceedings of the IEEE , volume=

    Gradient-based learning applied to document recognition , author=. Proceedings of the IEEE , volume=. 2002 , publisher=

    2002

  8. [8]

    Design and applications , volume=

    Recurrent neural networks , author=. Design and applications , volume=

  9. [9]

    Supervised sequence labelling with recurrent neural networks , pages=

    Long short-term memory , author=. Supervised sequence labelling with recurrent neural networks , pages=. 2012 , publisher=

    2012

  10. [10]

    3rd International Conference on Learning Representations (ICLR 2015) , year=

    Very deep convolutional networks for large-scale image recognition , author=. 3rd International Conference on Learning Representations (ICLR 2015) , year=

    2015

  11. [11]

    Advances in neural information processing systems , volume=

    Attention is all you need , author=. Advances in neural information processing systems , volume=

  12. [12]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , pages=

    2019

  13. [13]

    Visioncortex , title=

  14. [14]

    Potrace: a polygon-based tracing algorithm , author=

  15. [15]

    2008 , publisher=

    The art of artificial evolution: A handbook on evolutionary art and music , author=. 2008 , publisher=

    2008

  16. [16]

    IEEE Transactions on Visualization & Computer Graphics , volume=

    ClipGen: A Deep Generative Model for Clipart Vectorization and Synthesis , author=. IEEE Transactions on Visualization & Computer Graphics , volume=. 2022 , publisher=

    2022

  17. [17]

    Artificial Intelligence in Music, Sound, Art and Design , pages=

    Modern evolution strategies for creativity: Fitting concrete images and abstract concepts , author=. Artificial Intelligence in Music, Sound, Art and Design , pages=. 2022 , organization=

    2022

  18. [18]

    Proceedings of the 17th annual conference on Computer graphics and interactive techniques , pages=

    Paint by numbers: Abstract image representations , author=. Proceedings of the 17th annual conference on Computer graphics and interactive techniques , pages=

  19. [19]

    Workshops on Applications of Evolutionary Computation , pages=

    Genetic paint: A search for salient paintings , author=. Workshops on Applications of Evolutionary Computation , pages=. 2005 , organization=

    2005

  20. [20]

    IEEE Transactions on Visualization and Computer Graphics , volume=

    Abstract art by shape classification , author=. IEEE Transactions on Visualization and Computer Graphics , volume=. 2013 , publisher=

    2013

  21. [21]

    Roger Johansson , title=

  22. [22]

    Proceedings of the 25th annual conference on Computer graphics and interactive techniques , pages=

    Painterly rendering with curved brush strokes of multiple sizes , author=. Proceedings of the 25th annual conference on Computer graphics and interactive techniques , pages=

  23. [23]

    Salimans, J

    Evolution strategies as a scalable alternative to reinforcement learning , author=. arXiv preprint arXiv:1703.03864 , year=

    work page arXiv

  24. [24]

    ACM Transactions on Graphics (ToG) , volume=

    Modular primitives for high-performance differentiable rendering , author=. ACM Transactions on Graphics (ToG) , volume=. 2020 , publisher=

    2020

  25. [25]

    International Conference on Computational Intelligence in Music, Sound, Art and Design (Part of EvoStar) , pages=

    Paintings, polygons and plant propagation , author=. International Conference on Computational Intelligence in Music, Sound, Art and Design (Part of EvoStar) , pages=. 2019 , organization=

    2019

  26. [26]

    2022 , publisher=

    Computer vision: algorithms and applications , author=. 2022 , publisher=

    2022

  27. [27]

    2009 IEEE conference on computer vision and pattern recognition , pages=

    Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

    2009

  28. [28]

    Advances in neural information processing systems , volume=

    Laion-5b: An open large-scale dataset for training next generation image-text models , author=. Advances in neural information processing systems , volume=

  29. [29]

    International Conference on Learning Representations (ICLR) , year=

    A Neural Representation of Sketch Drawings , author=. International Conference on Learning Representations (ICLR) , year=

  30. [30]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    A learned representation for scalable vector graphics , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  31. [31]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Deepsvg: A hierarchical generative network for vector graphics animation , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  32. [32]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Im2vec: Synthesizing vector graphics without vector supervision , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  33. [33]

    SVGFusion: A VAE-Diffusion Transformer for Vector Graphic Generation

    SVGFusion: Scalable Text-to-SVG Generation via Vector Space Diffusion , author=. arXiv preprint arXiv:2412.10437 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  34. [34]

    ACM Transactions on Graphics (TOG) , numpages=

    Wang, Yizhi and Lian, Zhouhui , title=. ACM Transactions on Graphics (TOG) , numpages=

  35. [35]

    ACM Transactions on Graphics (TOG) , volume=

    Iconshop: Text-guided vector icon synthesis with autoregressive transformers , author=. ACM Transactions on Graphics (TOG) , volume=. 2023 , publisher=

    2023

  36. [36]

    International Conference on Machine Learning , pages=

    StrokeNUWA: Tokenizing Strokes for Vector Graphic Synthesis , author=. International Conference on Machine Learning , pages=. 2024 , organization=

    2024

  37. [37]

    ACM Transactions on Graphics (TOG) , volume=

    Differentiable vector graphics rasterization for editing and learning , author=. ACM Transactions on Graphics (TOG) , volume=. 2020 , publisher=

    2020

  38. [38]

    International Conference on Machine Learning (ICML) , pages=

    Learning transferable visual models from natural language supervision , author=. International Conference on Machine Learning (ICML) , pages=. 2021 , organization=

    2021

  39. [39]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Clipascene: Scene sketching with different types and levels of abstraction , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  40. [40]

    ACM Transactions on Graphics (TOG) , volume=

    Word-as-image for semantic typography , author=. ACM Transactions on Graphics (TOG) , volume=. 2023 , publisher=

    2023

  41. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Vectorfusion: Text-to-svg by abstracting pixel-based diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  42. [42]

    Advances in Neural Information Processing Systems , volume=

    Diffsketcher: Text guided vector sketch synthesis through latent diffusion models , author=. Advances in Neural Information Processing Systems , volume=

  43. [43]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    NIVeL: Neural Implicit Vector Layers for Text-to-Vector Generation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  44. [44]

    ACM Transactions on Graphics (TOG) , volume=

    Text-to-vector generation with neural path representation , author=. ACM Transactions on Graphics (TOG) , volume=. 2024 , publisher=

    2024

  45. [45]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    SVGDreamer: Text guided svg generation with diffusion model , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  46. [46]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Svgdreamer++: Advancing editability and diversity in text-guided svg generation , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  47. [47]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Sketchagent: Language-driven sequential sketch generation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  48. [48]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Towards layer-wise image vectorization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  49. [49]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Segment anything , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  50. [50]

    ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    SAMVG: A multi-stage image vectorization model with the segment-anything model , author=. ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2024 , organization=

    2024

  51. [51]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    SLIC superpixels compared to state-of-the-art superpixel methods , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2012 , publisher=

    2012

  52. [52]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Supersvg: Superpixel-based scalable vector graphics synthesis , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  53. [53]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Optimize & reduce: a top-down approach for image vectorization , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  54. [54]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Less is More: Efficient Image Vectorization with Adaptive Parameterization , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  55. [55]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Layered image vectorization via semantic simplification , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  56. [56]

    Proceedings of the SIGGRAPH Asia 2025 Conference Papers , year =

    LayerPeeler: Autoregressive Peeling for Layer-wise Image Vectorization , author =. Proceedings of the SIGGRAPH Asia 2025 Conference Papers , year =

    2025

  57. [57]

    arXiv preprint arXiv:2511.17454 , year=

    Illustrator's Depth: Monocular Layer Index Prediction for Image Decomposition , author=. arXiv preprint arXiv:2511.17454 , year=

    work page arXiv

  58. [58]

    Layertracer: Cognitive-aligned layered svg synthesis via diffusion transformer.arXiv preprint arXiv:2502.01105, 2025

    Layertracer: Cognitive-aligned layered svg synthesis via diffusion transformer , author=. arXiv preprint arXiv:2502.01105 , year=

    work page arXiv

  59. [59]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Empowering llms to understand and generate complex vector graphics , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  60. [60]

    ACM Transactions on Graphics , volume=

    Joint stroke tracing and correspondence for 2d animation , author=. ACM Transactions on Graphics , volume=. 2024 , publisher=

    2024

  61. [61]

    Le and Ruslan Salakhutdinov , title =

    Zihang Dai and Zhilin Yang and Yiming Yang and Jaime Carbonell and Quoc V. Le and Ruslan Salakhutdinov , title =. 2019 , note =

    2019

  62. [62]

    CVPR , year =

    Shenghai Yuan and Jinfa Huang and Xianyi He and Yunyuan Ge and Yujun Shi and Liuhan Chen and Jiebo Luo and Li Yuan , title =. CVPR , year =

  63. [63]

    Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution , booktitle =

    Zhou,. Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video Super-Resolution , booktitle =. 2024 , howpublished =

    2024

  64. [64]

    Differentiable Vector Graphics Rasterization for Editing and Learning , year =

    Tzu-Mao Li and Michal Luk. Differentiable Vector Graphics Rasterization for Editing and Learning , year =

  65. [65]

    CSS Transforms Module Level 1 , year =

  66. [66]

    SVG Animations , year =

  67. [67]

    SVG attribute: d , year =

  68. [68]

    Document Object Model (DOM) Specifications , year =

  69. [69]

    GPT-5.2 Model | OpenAI API , year =

  70. [70]

    Gemini 3 Pro (Vertex AI Generative AI models documentation) , year =

  71. [71]

    Peters and Arman Cohan , title =

    Iz Beltagy and Matthew E. Peters and Arman Cohan , title =. 2020 , note =

    2020

  72. [72]

    2024 , note =

    Tiffany Tseng and Ruijia Cheng and Jeffrey Nichols , title =. 2024 , note =

    2024

  73. [73]

    2013 , publisher =

    Dan Saffer , title =. 2013 , publisher =

    2013

  74. [74]

    Qwen3 Technical Report

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

    work page internal anchor Pith review Pith/arXiv arXiv

  75. [77]

    Training language models to follow instructions with human feedback

    Training Language Models to Follow Instructions with Human Feedback , author =. arXiv preprint arXiv:2203.02155 , year =

    work page internal anchor Pith review arXiv

  76. [79]

    Perception Models: Powerful Models for Image, Video, and Audio Perception , year =

  77. [81]

    Qwen/Qwen3-VL-8B-Thinking (Model Card) , author =

  78. [82]

    Playwright Documentation , author =

  79. [84]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics: System Demonstrations , year =

    LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models , author =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics: System Demonstrations , year =

  80. [86]

    SC20: International Conference for High Performance Computing, Networking, Storage and Analysis , year =

    ZeRO: Memory Optimizations Toward Training Trillion Parameter Models , author =. SC20: International Conference for High Performance Computing, Networking, Storage and Analysis , year =

Showing first 80 references.