Abstract
Animating Scalable Vector Graphics (SVGs) typically requires manual, labor-intensive processes such as rigging and keyframing. While text-to-video (T2V) diffusion models offer powerful motion synthesis, they operate in the pixel domain, creating a significant gap when applied to vector paths. We present LINR Bridge, a framework that utilizes Layered Implicit Neural Representations (LINR) as an intermediary to bridge the gap between mathematical vector paths and pixel-based video diffusion priors. By optimizing a neural representation through Video Score Distillation Sampling (VSDS), our method generates smooth, text-driven animations that preserve the structural integrity and aesthetic style of the original SVG.
Introduction
Vector graphics are essential in modern design due to their infinite scalability and editability. However, their animation remains a bottleneck. Existing automated approaches often suffer from two extremes: direct optimization of SVG points leads to "structural melting" and distortion, while rigid skeleton-based methods result in stiff, robotic movements. Furthermore, standard video models frequently fail to maintain the specific color palettes and flat-shaded aesthetics characteristic of vector art. LINR Bridge addresses these challenges by translating SVGs into a layered neural space where motion can be synthesized without compromising the underlying geometry.
Methodology
The LINR Bridge framework operates in three distinct stages to ensure high-fidelity motion transfer from diffusion models to vector graphics.
1. Vector Graphics Reconstruction
The input SVG is first converted into a Layered Implicit Neural Representation (LINR). Unlike standard coordinate-based networks, this model maps spatial coordinates to specific color intensities based on the SVG's individual layers. This architectural choice preserves sharp boundaries and the "flat" aesthetic of the original graphic during the optimization process.
2. Coarse Animation Generation
The reconstructed LINR is replicated to form a static video sequence. We then employ Video Score Distillation Sampling (VSDS) to extract motion knowledge from a pretrained T2V diffusion model. To prevent the animation from drifting away from the original design (identity drift), a Stability Extension is introduced, anchoring the generated frames to the initial static design.
3. Animation Refinement
The synthesized motion in the neural domain is transferred back to the original SVG. This is achieved by either calculating optical flow to warp the SVG's control points or optimizing the SVG parameters against the LINR frames using a structural loss function to ensure clean, mathematical paths.
Experimental Results
We evaluated LINR Bridge on a dataset of 80 SVGs across categories including humans, animals, and icons. Our method demonstrates significant improvements over state-of-the-art tools like LiveSketch and AniClipart in both quantitative metrics and qualitative fidelity.
Quantitative Performance
- Appearance Consistency: LINR Bridge achieved a CLIP cosine similarity score of 0.9649, outperforming LiveSketch (0.8618) and AniClipart (0.9476).
- Motion-Prompt Alignment: Our framework reached an X-CLIP score of 21.5139, indicating that the generated animations (e.g., a butterfly flapping its wings) align more closely with text prompts than competing methods.
Qualitative Analysis
Visual comparisons reveal that while LiveSketch often causes shapes to distort and AniClipart produces rigid, unnatural motion, LINR Bridge generates fluid, organic movements. The use of a Binarization Loss further ensures that edges remain crisp, mimicking the precision of hand-crafted vector animations.
Discussion and Conclusion
The success of LINR Bridge stems from the "Neural Bridge" concept: Implicit Neural Representations share the infinite resolution of SVGs while being differentiable and compatible with pixel-based optimization. While the current approach excels at flat-shaded graphics, future work will focus on representing more complex SVG features such as gradient meshes and semi-transparent overlapping layers. Despite the computational cost associated with VSDS optimization, LINR Bridge represents a significant step toward high-quality, automated vector animation.
Summary updated