Pengyu965/StructHub

Back2Struct: Making Structured Images Editable Again

Recovering editable, object-level SVG / XML code from images of structured graphics.

• 📄 Paper: arXiv:2609.37016 — Back2Struct: Making Structured Images Editable Again
• 🔗 Project page: https://pengyu965.github.io/Back2Struct.github.io/
• 🤖 Model: Pengyu965/Back2Struct-Image2SVG-7B
• 📚 Dataset: Pengyu965/StructHub

Back2Struct "makes structured images editable again" by directly recovering vector graphics code (SVG / XML) from image representations. Given an image of a structured graphic — a diagram, chart, flowchart, architecture figure, UML diagram, or schema — it predicts semantically object-level SVG/XML that explicitly encodes text, shapes, topology, and layout (not low-level pixel tracing), so the code imports straight into tools like PowerPoint to edit, restyle, and reuse while preserving structure. Training is supervised fine-tuning from Qwen2.5-VL-7B-Instruct, followed by GRPO with a composite reward (syntactic validity, length fidelity, perceptual quality).

This project ships two artifacts — a model and a dataset — documented together in this shared card.

🤖 Model — Back2Struct-Image2SVG-7B

Base	Qwen/Qwen2.5-VL-7B-Instruct (7B)
Stage 1 — SFT	Supervised fine-tuning on StructHub image→SVG pairs
Stage 2 — GRPO	Group-Relative Policy Optimization (TRL)
Reward	Composite, render-gated: syntactic validity (must render) × perceptual fidelity (baseline-subtracted DINOv2 similarity between the rendered prediction and the target), encouraging compilability, length consistency, and structural/semantic faithfulness
Max output	8,192 tokens · LoRA (r=64, α=128), merged into the released weights

📚 Dataset — StructHub

Structured images paired with their editable SVG/XML source code.

Split	Description	# examples
train	Full training set (all sources, all token lengths), benchmark held out	86,102
benchmark	Held-out evaluation benchmark (starvector 569 + crawled 334 + nn_diagram 97)	1,000

Fields: id (string), source (string), token_count (int32), image (embedded PNG), svg (ground-truth SVG/XML). The benchmark is stratified by SVG length into easy (≤2,048), medium (2,048–4,096), and hard (>4,096) tiers (333/333/334).

Results

StructHub benchmark, render-gated (a non-rendering prediction scores 0 on DINO/SSIM/GPT, 1 on LPIPS). SR = render success (%); GPT = 0–100 GPTScore.

Model	SR ↑	DINO ↑	LPIPS ↓	SSIM ↑	GPT ↑
Overall
Qwen3-VL-30B-A3B	18.8	0.1566	0.9067	0.1201	13.29
StarVector-8B	31.4	0.2694	0.7936	0.2170	21.05
Qwen2.5-VL-7B (base)	52.3	0.3527	0.7834	0.3276	27.56
Qwen2.5-VL-7B (SFT)	42.8	0.3715	0.7403	0.2891	33.49
Back2Struct (RL)	71.5	0.5622	0.6473	0.4966	51.85
Easy
Qwen3-VL-30B-A3B	26.5	0.2231	0.8659	0.1702	19.51
StarVector-8B	46.1	0.3923	0.7023	0.3123	31.90
Qwen2.5-VL-7B (base)	69.3	0.4776	0.7123	0.4196	39.66
Qwen2.5-VL-7B (SFT)	56.0	0.4866	0.6630	0.3789	44.58
Back2Struct (RL)	84.0	0.6384	0.5956	0.5796	62.63
Medium
Qwen3-VL-30B-A3B	19.6	0.1631	0.9004	0.1288	13.38
StarVector-8B	37.0	0.3232	0.7533	0.2583	25.27
Qwen2.5-VL-7B (base)	51.8	0.3546	0.7816	0.3306	26.88
Qwen2.5-VL-7B (SFT)	45.5	0.3985	0.7161	0.3091	36.50
Back2Struct (RL)	75.6	0.6088	0.6182	0.5275	55.14
Hard
Qwen3-VL-30B-A3B	10.2	0.0836	0.9536	0.0614	7.01
StarVector-8B	11.1	0.0932	0.9249	0.0808	6.02
Qwen2.5-VL-7B (base)	35.7	0.2263	0.8561	0.2331	16.17
Qwen2.5-VL-7B (SFT)	27.0	0.2297	0.8413	0.1795	19.43
Back2Struct (RL)	55.0	0.4399	0.7278	0.3830	37.83

On successfully compiled outputs, the 7B model matches or beats frontier closed models:

Model	DINO ↑	LPIPS ↓	SSIM ↑	GPT ↑
Gemini-2.5-Pro	0.8855	0.4648	0.6532	77.95
GPT-5	0.8852	0.4350	0.6573	77.21
Back2Struct (7B)	0.8697	0.3896	0.6738	78.63

Usage

Run the model

import torch
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info

model_id = "Pengyu965/Back2Struct-Image2SVG-7B"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)

messages = [{"role": "user", "content": [
    {"type": "image", "image": "diagram.png"},
    {"type": "text",  "text": "Convert this structured image into editable SVG/XML code."},
]}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(text=[text], images=image_inputs, videos=video_inputs,
                   padding=True, return_tensors="pt").to(model.device)
generated = model.generate(**inputs, max_new_tokens=8192, do_sample=False)
print(processor.batch_decode(generated[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0])

Load the dataset

from datasets import load_dataset
ds = load_dataset("Pengyu965/StructHub")
ex = ds["benchmark"][0]
ex["image"].save("input.png"); print(ex["svg"][:500])

Intended use & limitations

• Intended use: recovering editable vector code from images of structured graphics.
• Not for: photographic / natural-image vectorization.
• Limitations: built on a general (non-code-specialized) VLM; outputs capped at 8,192 tokens; very long/dense diagrams (hard tier) remain the main failure mode (render reliability).

Licensing

Model weights are released under Apache-2.0 (inheriting the Qwen2.5-VL-7B-Instruct base). The StructHub dataset aggregates multiple sources, including crawled open-domain graphics; it is released for research use — review and respect the original content licenses. (Confirm before wide distribution.)

Citation

@misc{yan2026back2structmakingstructuredimages,
      title={Back2Struct: Making Structured Images Editable Again},
      author={Pengyu Yan and Yixin Wu and Yunjie Tian and David Doermann},
      year={2026},
      eprint={2609.37016},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.37016},
}