Research map / Vision-language and multimodal / Human and embodied modeling

OmniSVG: A Unified Scalable Vector Graphics Generation Model

NeurIPS 2025 · Vision-language and multimodal › Human and embodied modeling · approach: Medical vision-language models · 5 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.

Builds OmniSVG, a Qwen2.5-VL-based autoregressive model that generates colored SVGs from text, image, and character-reference prompts using parameterized SVG command/coordinate tokens, and introduces the MMSVG-2M dataset and MMSVG-Bench.

Problem

Generating high-quality, complex, editable SVGs under multimodal instructions; existing optimization-based methods are slow and produce redundant paths, while autoregressive methods are limited to simple monochrome icons and short context lengths.

Approach

Parameterize SVG commands, coordinates (merged into one token via x*w+y), and fill colors into discrete tokens and fine-tune a pre-trained VLM (Qwen2.5-VL) with next-token prediction, enabling sequences over 30k tokens and multimodal conditioning.

Key results

• On MMSVG-Icon text-to-SVG, OmniSVG-L(8B) reaches FID 130.56 vs. 213.28 for IconShop and 250.77 for VectorFusion, using 5.7k tokens.
• On MMSVG-Icon image-to-SVG, OmniSVG(4B) reaches DINO 0.993 and SSIM 0.950, matching VTracer's DINO (0.993) but with lower SSIM (VTracer 0.966) and fewer tokens than DiffVG (322.0k).
• Parameterizing both coordinates and colors reduces average tokens from 18.5k to 4.8k in the ablation while improving FID from 218.76 to 145.89.

Measured effects reported in the paper

Surprising findings

• VTracer, a non-learned tracing algorithm, achieves the best image-to-SVG DINO, SSIM and LPIPS on icons and illustrations, outperforming OmniSVG on those metrics.
• The 7B Qwen2.5-VL variant gets a better FID (134.45) than the 3B variant (145.89) but lower CLIP score (0.254 vs 0.308) in the VLM ablation.

Limitations

• Generates tens of thousands of tokens for complex samples, causing considerable generation time.
• Only bounded to vector-style image prompts and fails on natural images.
• Vtracer and DiffVG achieve better image reconstruction metrics on several subsets, so the image-to-SVG advantage is not uniform.

Experimental setup

Datasets: MMSVG-2M FIGR-8-SVG IconShop dataset

Benchmarks: MMSVG-Bench

Metrics: FID CLIP score Aesthetic score HPS DINO score SSIM LPIPS MSE GPT-4o score for character reference token length

Baselines: VectorFusion SVGDreamer Chat2SVG IconShop LIVE DiffVG GPT-4o StarVector VTracer

Builds on: Qwen2.5-VL IconShop

Closest papers by meaning

• Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models (CVPR 2025 · similarity 0.91)
• SVGBuilder: Component-Based Colored SVG Generation with Text-Guided Autoregressive Transformers (AAAI 2025 · similarity 0.90)
• VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models (NeurIPS 2026 · similarity 0.89)
• SVGDreamer: Text Guided SVG Generation with Diffusion Model (CVPR 2024 · similarity 0.89)
• StarVector: Generating Scalable Vector Graphics Code from Images and Text (AAAI 2025 · similarity 0.89)
• InternSVG: Towards Unified SVG Tasks with Multimodal Large Language Models (ICLR 2026 · similarity 0.89)