Research map / Vision-language and multimodal / Multimodal LLMs

VectorArk: Learning Practical Image Vectorization with Rounded Polygon Representation

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Fine-tunes InternVL2-1B to map degraded outline rasters to a compact rounded-polygon SVG representation, with a vectorization-based degradation model and DINO-ranked test-time scaling.

Problem

VLM-based image vectorization is evaluated on clean synthetic rasters and generalizes poorly to real inputs such as text-to-image outputs or different rasterizers; autoregressive SVG generation is also unstable across samples.

Approach

Represent SVG paths as rounded polygons (vertices plus per-vertex roundness), feed the VLM a decolorized outline raster produced by a classical vectorizer, train on classically vectorized degraded renders, and select the best of N samples by DINO similarity to the input.

Key results

• On SVGenius (Hard), VectorArk reaches SSIM 0.830, LPIPS 0.120, DINO 0.958 versus StarVector 8B's 0.672, 0.258, 0.893.
• Rounded-polygon tokens are 27.9% to 46.6% shorter than OmniSVG's representation while keeping DINO above 0.99.
• Test-time scaling on SVGenius raises DINO from 0.813 (n=1) to 0.905 (n=64), an improvement of +11.2%.

Surprising findings

• LoRA and LLM-only finetuning performed worse than full finetuning including the ViT encoder.
• Training on color rasters underperforms outline-only inputs, with the gap largest on hard samples.

Limitations

• Simplifies highly detailed illustrations, text and gradient effects; dense local path structure remains challenging.
• Post-processing color, z-order and stroke recovery cannot handle gradients, filters or transparency.
• Relies on a classical vectorizer for the input outline, which can be slow (LIVE takes over 10 minutes per image).

Experimental setup

Datasets: ~5M SVGs from icons, logos, and flat graphics SArena SVGenius SVG-Stack SVG-Fonts SVG-Icons SVG-Emoji MMSVG-Icon SVGX core 250k

Benchmarks: SArena SVGenius SVG-Stack SVG-Fonts SVG-Icons SVG-Emoji MMSVG-Icon SVGX core 250k

Metrics: SSIM LPIPS MSE DINO score Chamfer distance

Baselines: GPT-4o Gemini-2.5-Pro OmniSVG StarVector 8B

Builds on: InternVL2 StarVector OmniSVG DiffVG Adobe Illustrator Image Trace

Closest papers by meaning

• Optimize & Reduce: A Top-Down Approach for Image Vectorization (AAAI 2024 · similarity 0.87)
• VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models (NeurIPS 2026 · similarity 0.86)
• Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models (CVPR 2025 · similarity 0.85)
• Layered Image Vectorization via Semantic Simplification (CVPR 2025 · similarity 0.84)
• Channel-wise Vector Quantization (NeurIPS 2026 · similarity 0.84)
• Less is More: Efficient Image Vectorization with Adaptive Parameterization (CVPR 2025 · similarity 0.84)