Pengyu965/Back2Struct-Image2SVG-7B

VISIONPricing:Input $0.4 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:32kPublished:Sep 29, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Pengyu965/Back2Struct-Image2SVG-7B is a 7 billion parameter vision-language model developed by Pengyu Yan, Yixin Wu, Yunjie Tian, and David Doermann, fine-tuned from Qwen2.5-VL-7B-Instruct. This model specializes in converting images of structured graphics into editable, object-level SVG/XML code, preserving text, shapes, topology, and layout. It is optimized for recovering vector graphics from diagrams, charts, and flowcharts, enabling direct import into editing tools like PowerPoint. The model achieves high render success and perceptual fidelity on the StructHub benchmark, outperforming larger frontier models on successfully compiled outputs.

Loading preview...

Back2Struct: Image to Editable SVG/XML

Back2Struct-Image2SVG-7B is a 7 billion parameter vision-language model designed to transform images of structured graphics into editable SVG/XML code. Developed by Pengyu Yan et al., this model is fine-tuned from Qwen2.5-VL-7B-Instruct and utilizes a novel Group-Relative Policy Optimization (GRPO) with a composite reward function to ensure syntactic validity, length fidelity, and perceptual quality of the generated code.

Key Capabilities

  • Object-level Vector Recovery: Unlike pixel tracing, it generates semantically meaningful SVG/XML that explicitly encodes text, shapes, topology, and layout.
  • Direct Editability: The output code can be directly imported into tools like PowerPoint for editing, restyling, and reuse while maintaining structural integrity.
  • High Fidelity: Achieves a render success rate of 71.5% overall on the StructHub benchmark, significantly outperforming base models and even larger frontier models like Qwen3-VL-30B-A3B and StarVector-8B.
  • Benchmark Performance: On successfully compiled outputs, the 7B model matches or surpasses frontier closed models like Gemini-2.5-Pro and GPT-5 in perceptual metrics (DINO, LPIPS, SSIM) and GPTScore.

Good For

  • Converting diagrams, charts, flowcharts, architecture figures, and UML diagrams from image format to editable vector graphics.
  • Automating the process of making structured visual content reusable and modifiable.

Limitations

  • Not intended for photographic or natural-image vectorization.
  • Output is capped at 8,192 tokens, which can be a limitation for very long or dense diagrams.
  • Performance on the 'hard' tier (very long/dense diagrams) remains the primary area for improvement in render reliability.

Finetunes
1 models