Research map / Vision-language and multimodal / Multimodal LLMs

InternSVG: Towards Unified SVG Tasks with Multimodal Large Language Models

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants

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.

Introduces the InternSVG family (SAgoge dataset of ~16M SVG samples, SArena benchmark, and InternSVG, a unified InternVL3-8B-based MLLM) for SVG understanding, editing, and generation, trained with a two-stage curriculum and subword-initialized SVG special tokens.

Problem

SVG modeling is fragmented across single-task datasets and benchmarks, with small scale, limited transferability across understanding/editing/generation, and difficulty handling long, complex SVG code.

Approach

A single MLLM is trained jointly on understanding, editing, and generation for SVG code, using SVG-specific special tokens (tags, attributes, numbers) initialized as averages of their subword embeddings, plus a curriculum from short icons/chemistry SVGs to long illustrations and animations.

Key results

• On SArena-Icon understanding, InternSVG 8B reaches 85.1 overall accuracy vs 77.1 for Claude-Sonnet-4 (+8 points).
• On SArena-Icon Text-to-SVG, FID drops to 8.715 from 15.840 for Claude-Sonnet-4 (about 45% lower, reported as 56% in intro).
• Two-stage training reduces FID-C on SArena-Illustration Text-to-SVG from 25.671 (one-stage) to 5.141 and raises DINO from 0.830 to 0.924 for Image-to-SVG.

Surprising findings

• Multi-task joint training (understanding, editing, generation) improves generation metrics over single-task training, suggesting positive cross-task transfer.
• Several large open-source MLLMs (e.g., InternVL3.5 variants) show poor Text-to-SVG on long illustrations, with very high token counts and low success rates.

Limitations

• Style transfer editing still lags Claude-Sonnet-4 on SArena-Icon.
• Chemistry Text-to-SVG CLIP-T2I and some generation metrics are not uniformly best; Text-to-SVG CLIP-T2I trails Claude-Sonnet-4 on icons.
• Evaluation uses a black-image penalty for unrenderable SVGs, which affects metric interpretation.

Experimental setup

Datasets: SAgoge Iconfont SVGRepo OpenClipart PubChem SVG-Stack UniSVG SGP-Bench

Benchmarks: SArena-Icon SArena-Illustration SArena-Chemistry SArena-Animation SGP-Bench SVG-Stack UniSVG

Metrics: accuracy DINO SSIM LPIPS PSNR FID FID-CLIP CLIP-T2I CLIP-I2I FVD

Baselines: GPT-4o Claude-Sonnet-4 Gemini-2.5-Flash Qwen2.5-VL-72B InternVL3-78B Llama-4-Maverick LLM4SVG StarVector OmniSVG IconShop VectorFusion SVGDreamer

Builds on: InternVL3 Qwen2.5 DiffVG LIVE StarVector

Closest papers by meaning

• Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models (CVPR 2025 · similarity 0.90)
• Empowering LLMs to Understand and Generate Complex Vector Graphics (CVPR 2025 · similarity 0.90)
• VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models (NeurIPS 2026 · similarity 0.89)
• OmniSVG: A Unified Scalable Vector Graphics Generation Model (NeurIPS 2025 · similarity 0.89)
• SVGBuilder: Component-Based Colored SVG Generation with Text-Guided Autoregressive Transformers (AAAI 2025 · similarity 0.88)
• SVGDreamer: Text Guided SVG Generation with Diffusion Model (CVPR 2024 · similarity 0.87)