PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2605.01517v1cs.CV

TL;DR

Text-to-SVG animation must bridge discrete code generation and continuous visual dynamics while avoiding context growth, identity drift, and rigid motion. VAnim uses Sparse State Updates on a persistent SVG DOM, identification-first planning, and rendering-aware GRPO, alongside the SVGAnim-134k benchmark. The paper reports state-of-the-art semantic alignment and structural validity, with additional evaluations of non-rigid deformation, temporal smoothness, identity preservation, and motion magnitude.

  • Problem

    Text-to-SVG animation is limited by context explosion, identity drift, weak supervision of rendered motion, and rigid affine-transform biases.

  • Method

    VAnim combines Sparse State Updates on a persistent SVG DOM, Identification-First Motion Planning, and rendering-aware GRPO trained with visual feedback.

  • Results

    VAnim achieves state-of-the-art semantic alignment and structural validity and supports complex path-level non-rigid deformations in open-domain vector animation.

  • Takeaways & Limitations

    The work establishes SVGAnim-134k as a benchmark and presents structure-preserving vector animation with supplementary measures for temporal smoothness, identity preservation, and motion magnitude.

  • Takeaways & Limitations

    Rendering-aware reinforcement learning adds computational overhead, and generalization beyond Lottie-style vector graphics remains open.

Abstract

from arXiv · show

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.

1. Introduction

Text-to-SVG animation is hindered by long-horizon context growth, identity drift, limited non-rigid motion, and insufficient structured data. VAnim addresses these issues with sparse DOM updates, identification-first planning, rendering-aware optimization, and SVGAnim-134k.

  • Challenges: Frame-by-frame SVG generation causes context explosion and identity drift because verbose documents and static attributes are repeatedly regenerated.These effects can exceed context windows and make object identities flicker or collapse over time.
  • Limitations of Existing Methods: Existing optimization-based methods can be slow and topologically unstable, while CSS/SMIL-oriented systems are restricted to affine motion.These limitations hinder professional designs with closed shapes, occlusions, and non-rigid behaviors.
  • VAnim: VAnim models animation as Sparse State Updates on a persistent SVG DOM, reducing effective sequence length by more than 9× while preserving structure and inactive elements.The method updates selected identifier-anchored attributes instead of rewriting the full SVG at every timestep.
  • VAnim: Identification-First Motion Planning links semantic entities to SVG IDs and plans temporal behavior before emitting code edits.This grounds motion in persistent structural components and reduces structural dissociation or identity loss.
  • VAnim: Rendering-aware reinforcement learning with GRPO uses video-perception feedback to improve semantic alignment and support geometry-level non-rigid deformations.The approach directly addresses the limited rendered-motion supervision provided by SFT.
  • Dataset: SVGAnim-134k establishes a large-scale benchmark of professionally authored vector animations with structured semantics and hierarchical grouping.Its construction includes structural standardization, sparse encoding, dual-stream annotation, and ID-consistency filtering.

2. SVGAnim-134k: A Large-Scale Benchmark for Vector Animation

SVGAnim-134k is a large-scale vector-animation benchmark built from structurally rigorous, professionally authored animations. Its sparse state representation preserves persistent SVG topology while reducing sequence redundancy and supporting structure-grounded supervision.

  • Dataset construction: The dataset pipeline includes acquisition, topological canonicalization, sparse state extraction, and dual-stream annotation generation.Lottie animation files are converted into explicit SVG sequences with persistent identifiers for dynamically updated nodes.
  • Representation: Animation sequences are represented as topologically isomorphic SVG DOM trees whose geometry, transformation, and appearance attributes evolve over time.This attribute-driven representation provides the structural basis for sparse state modeling.
  • Sparse state modeling: Sparse updates encode only changed attributes for persistent node IDs, preserving static elements and anchoring dynamic changes to the SVG DOM.The complete animation is represented as (S0, ∆1, . . . , ∆T), while serialized control tags identify time steps and nodes.
  • Sparse state modeling: 9.86× compression reduces a 24-frame animation from 86.0k to 9.2k tokens using sparse state updates.The representation stores an initial SVG and attribute differentials, with the diff component comprising 61% of the sequence length.
  • Annotation and supervision: Structure-Bound Chain-of-Thought supervision links user prompts to visual entities and temporal dynamics through explicit SVG IDs.The strict ID-consistency filter discards chains referencing nonexistent IDs or misidentified node types.
  • Framework integration: The framework overview combines Identification-First Motion Planning with Sparse State Updates and rendering-aware reinforcement learning.The two stages ground visual entities and update only targeted DOM nodes, while GRPO uses hybrid visual and structural rewards.

3. Methodology

VAnim formulates text-to-SVG animation as planned, identifier-grounded sparse updates to a persistent SVG structure. It combines supervised SVG learning with rendering-aware GRPO, using visual-semantic and format-validity rewards to support precise non-rigid motion.

  • Problem formulation: Given an initial SVG, its rendered image, and a natural-language instruction, VAnim generates sparse state updates across the animation frames.The output is D = {∆t | t = 1, . . . , T}.
  • Problem formulation: VAnim introduces a latent Structure-Bound Chain-of-Thought variable that separates semantic planning from sparse-update generation.The input combines visual context, SVG structure, and the prompt, while the output contains the reasoning trace and updates.
  • Model architecture: The multimodal architecture interleaves visual embeddings from the rendered SVG with textual embeddings from SVG code and user prompts.A visual encoder produces continuous embeddings that are projected into the language model’s token space.
  • Inference: Coarse-to-fine inference first identifies relevant entities and intended motions, then executes localized attribute updates constrained by the SVG DOM.This separation is designed to prevent identity drift and unbounded structural modification.
  • Inference: Identification-First Motion Planning grounds instructions in persistent SVG IDs and attributes before generating a structured motion blueprint.The Director role resolves what to move and where, restricting the subsequent generation space.
  • Inference: Sparse State Update Execution predicts token-level differentials rather than rewriting full SVG code, preserving non-participating elements and topology.The Animator role generates updates conditioned on the motion blueprint.
  • Training: SFT initializes SVG literacy and valid YAML-based diffs, while rendering-aware RL addresses the gap between syntactic correctness and perceptual motion quality.The RL stage uses the SVGAnim-RL subset and evaluates rendered candidate videos.
  • Training: The hybrid reward jointly optimizes semantic fidelity and format validity using PE-Core feedback and strict structural constraints.Validity requires parseable and renderable sequences, exact target length, and valid IDs without topology breakage.

4. Experiments

Experiments evaluate VAnim against language-model and optimization baselines using semantic alignment, executability, qualitative comparisons, and ablations. VAnim combines sparse state updates, visual grounding, and rendering-aware reinforcement learning to preserve structure while producing precise motion.

  • Qualitative Comparison: VAnim preserves visual identity and topology, whereas GPT-5.2 changes the toucan’s wing color, Gemini distorts the bottle, and LiveSketch produces broken lines.
  • Quantitative Comparison: VAnim-GRPO achieves a Semantic Alignment of 0.281 and a Success Rate of 100% on SVGAnim-Test.
  • Additional Ablations: The no-SSU baseline drops Success Rate to 62.3%, while removing input images reduces SSIM and motion quality.
  • Qualitative Comparison: Direct path-level editing enables non-rigid deformation beyond the affine transformations used by CSS/SMIL-based baselines.
  • Ablation Study: Removing Rendering-Aware RL reduces semantic alignment from 0.281 to 0.268 and produces conservative motion that under-satisfies requested deformation.
  • Ablation Study: Removing Structure-Bound CoT reduces semantic alignment from 0.281 to 0.255 and causes manipulation of incorrect SVG entities.

5. Conclusion

VAnim introduces an LLM-based framework and SVGAnim-134k dataset for open-domain text-to-SVG animation. Its sparse DOM updates, identification-first planning, and rendering-aware reinforcement learning support structure-preserving, semantically aligned deformation, while interactive behaviors and longer narratives remain future work.

  • VAnim is an LLM-based framework for open-domain text-to-SVG animation, accompanied by the SVGAnim-134k dataset.
  • Sparse State Updates on a persistent SVG DOM, together with Identification-First Motion Planning, preserve structure and non-participating elements.
  • Experiments report state-of-the-art semantic alignment and structural validity for complex path-level non-rigid deformations.
  • Extending the framework to interactive behaviors and longer, multiscene narratives remains future work.

Impact Statement

VAnim is intended for professional creative applications but may also be misused to create deceptive or manipulated content. Its rendering-aware reinforcement learning adds computational overhead, and its generalization beyond Lottie-style graphics remains open.

  • VAnim is intended as a creative assistant for professional designers and developers in design, UI/UX, content creation, and accessibility.
  • The system could potentially be misused to create deceptive or manipulated content, motivating responsible deployment and transparency.
  • Rendering-aware reinforcement learning creates computational overhead that may limit accessibility for resource-constrained users.
  • The approach is primarily designed for Lottie-style vector graphics, and generalization to hand-authored or tool-exported SVGs remains an open question.

Overview

The supplementary material organizes expanded evaluations, implementation details, user studies, qualitative results, related work, and limitations. It also reports an expanded evaluation matrix covering independent semantic, motion, smoothness, identity, and ablation analyses.

  • Supplementary Organization: Section A provides the full evaluation table with additional baselines, input and reward ablations, and GRPO sensitivity analysis.
  • Supplementary Organization: Section B covers implementation details, while Section C reports a user study comparing VAnim with existing approaches.
  • Supplementary Organization: Sections D and E provide additional qualitative results and related work, respectively.
  • Supplementary Organization: Section F discusses limitations and potential societal impact.
  • Expanded Evaluation: The expanded evaluation matrix adds InternVideo2, motion magnitude, temporal smoothness, identity preservation, and ablations isolating SSU, visual grounding, and reward terms.

A.1. Expanded Analysis

The expanded analysis isolates the roles of visual feedback, sparse updates, and training design choices in VAnim’s performance. Results show that removing SSU or specific reward inputs harms distinct aspects of animation quality, while larger GRPO groups trade improved alignment for lower SSIM.

  • Ablation findings: 62.3% success rate is reached by the no-SSU baseline, alongside much worse temporal smoothness than sparse-update modeling.Frame-by-frame regeneration preserves some visual quality on easy cases but performs poorly overall.
  • Ablation findings: Removing Ralign lowers semantic fidelity and motion magnitude, while removing Rfmt mainly reduces executability.The ablations separate the contributions of semantic and format rewards.
  • Ablation findings: Removing the input image degrades identity preservation and motion consistency.This identifies the input image as important for maintaining object identity during motion.
  • GRPO sweep: Larger GRPO group sizes improve semantic alignment and motion magnitude at the cost of slightly reduced SSIM.The authors retain G = 8 as the default trade-off between exploration and structural stability.
  • Training pipeline: SFT teaches SVG syntax, DOM awareness, structured reasoning, and sparse YAML-based diff prediction before reinforcement learning.The SFT checkpoint initializes subsequent GRPO training.
  • Rendering-aware RL: Rendering-aware reinforcement learning reconstructs sparse updates into full SVG sequences and renders short video clips for perceptual reward computation.Rendered videos are encoded with PE-Core-G14, with embeddings reused across reward terms.

C. User Study

The user study evaluates visual integrity, motion smoothness, and instruction following using blinded human ratings. VAnim receives the strongest reported results, especially for preserving structure and executing nuanced non-rigid motion.

  • Study setup: The study recruited 15 participants who each evaluated 15 randomly selected cases across four anonymized systems.Results were collected from professional UI designers and computer science researchers as well as other participants.
  • Study results: VAnim consistently outperforms all baselines across visual integrity, motion smoothness, and instruction following.The study uses mean ratings on a 1–5 Likert scale.
  • Visual Integrity: 4.62 is VAnim’s mean Visual Integrity score, compared with 2.15 for LiveSketch and 3.42 for Gemini 3 Pro.Visual integrity measures preservation of shape, identity, and topology without collapse or flicker.
  • Instruction Following: 4.55 is VAnim’s Instruction Following score, with participants noting successful execution of nuanced shape morphing.Gemini 3 Pro was described as often simplifying complex instructions into basic rigid transformations.
  • Dataset coverage: SVGAnim-134k spans UI components, organic entities, and complex grouped structures requiring rigid transitions, path deformation, and coordinated motion.Its samples include professional-grade topology, rich fills, and layered occlusion structures.
  • Dataset coverage: The dataset’s semantic coverage is presented as supporting open-domain vector motion rather than overfitting to one visual style.Complex samples include sub-components that move with parent structures without topological breakage.

D.2. Quantitative Analysis of Motion Primitives

SVGAnim-134k contains both rigid transformations and substantial path-level deformation, providing supervision for diverse motion primitives. The analysis compares these ground-truth distributions with model-generated motion patterns.

  • Ground-truth motion distribution: 70% of ground-truth state updates use rigid transformations, while 30% directly manipulate path data d.The path-level portion provides supervision for shape morphing and organic movement.
  • Ground-truth motion distribution: The 30% path-manipulation share confirms that SVGAnim-134k is not limited to simple affine motion.The benchmark therefore includes substantial non-rigid deformation complexity.
  • Distribution comparison: Figure S3 compares transform and d usage across original data and model generations.Alignment between generated and original distributions indicates how well models capture the dataset’s motion priors.
  • Relation to prior work: Optimization-based vector animation methods iteratively fit sketches or clipart under pretrained-model and differentiable-rendering supervision.The cited line of work includes LiveSketch, FlipSketch, and AniClipart.
  • Relation to prior work: Declarative CSS/SMIL approaches mainly operate on affine transforms and appearance attributes, preventing path-level deformation.This limitation motivates direct geometry updates for motions such as morphing and organic movement.

F. Limitations

VAnim’s limitations concern its structural assumptions, training domain, computational cost, and broader societal considerations. The method is most directly bounded by persistent topology and Lottie-derived data.

  • Structural scope: VAnim assumes a persistent SVG DOM tree with stable identifiers, making dynamic node insertion, deletion, and topology changes less suitable.This is a structural scope boundary of the current formulation.
  • Domain scope: Training and evaluation use a single Lottie-derived SVG source family, leaving generalization to hand-authored or design-tool-exported SVGs open.The limitation concerns out-of-domain generalization rather than the reported in-domain setting.
  • Computational scope: Rendering-aware reinforcement learning adds non-trivial compute because each update samples and renders multiple candidates per prompt.This may limit deployment in resource-constrained environments.
  • Societal impact: The paper directs societal-impact and ethical considerations to its separate Impact Statement.No further societal-impact claim is provided in the supplied limitation passages.
Loading 2605.01517v1…