VFIG-8B-SFT · Image-to-Code

Given an image of a scientific figure, this model generates the SVG source code that recreates it. The Stage 2 LoRA weights (rank 64, alpha 128) are merged into Qwen3-VL-8B-Instruct. The repository includes the BF16 model weights, tokenizer, image processor, and chat template.

import torch
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor

repo = "QijiaHe/VFIG-8B-SFT"
processor = AutoProcessor.from_pretrained(repo)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    repo, dtype=torch.bfloat16, device_map="auto"
).eval()

Paper · VFIG-Data · VFIG-4B after reinforcement learning