Parameter	Value
Base Model	Qwen3-VL-2B-Instruct
Dataset	ahazimeh/slide2svg (~39,500 samples)
Method	LoRA (r=64, alpha=128)
Epochs	2
Batch Size	16 (2 × 8 gradient accumulation)
Learning Rate	1e-4
Context Length	12,288 tokens
Precision	16-bit
Training Time	~5.5 hours
Hardware	NVIDIA RTX 5090 (32GB)

Evaluation Results

Metric	Score
Valid XML Structure	100% (20/20)
Has `<svg>` tag	100% (20/20)
Has `<foreignObject>`	100% (20/20)
Final Training Loss	0.091
Final Validation Loss	0.091
Avg Output Length	91% of reference

Usage

With Unsloth

python

from unsloth import FastVisionModel
from PIL import Image

# Load model
model, tokenizer = FastVisionModel.from_pretrained(
    "Bombek1/slide2svg-vl-2b",
    load_in_4bit=True,  # or False for 16-bit
)
FastVisionModel.for_inference(model)

# Prepare input
image = Image.open("your_slide.png")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": "Convert this presentation slide to SVG format."}
        ]
    }
]

# Generate
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
generated_ids = outputs[0][inputs["input_ids"].shape[1]:]
svg_output = tokenizer.decode(generated_ids, skip_special_tokens=True)

print(svg_output)

Output Format

The model generates SVG with the following structure:

xml

<?xml version="1.0" encoding="utf-8"?>
<html>
<svg xmlns="http://www.w3.org/2000/svg" width="1024" height="768" fill="white">
  <image x="0.0%" y="0.0%" width="100.0%" href="background.png" />
  <g id="images">
    <image x="14.3%" y="28.5%" width="35.8%" href="image_0.png" />
  </g>
  <g id="text">
    <foreignObject x="5.4%" y="8.1%" width="32.0%" height="12.0%" overflow="visible">
      <div xmlns="http://www.w3.org/1999/xhtml" style="font-family: Inter; font-size: 74px; font-weight: bold; color: #000000;">
        <div>Your Text Here</div>
      </div>
    </foreignObject>
  </g>
</svg>
</html>

Limitations

Outputs are ~91% the length of reference SVGs (may omit some minor elements)
Default output resolution is 1024×768 (positions use percentages, so they scale)
Best results on slides similar to the training data (presentation-style layouts)
Complex diagrams or charts may not be fully captured

License

Apache 2.0 (inherited from Qwen3-VL base model)

Acknowledgments

Base model: Qwen3-VL-2B-Instruct by Alibaba
Dataset: ahazimeh/slide2svg
Training framework: Unsloth