app.py

import gradio as gr
from transformers import AutoProcessor, VisionEncoderDecoderModel
from PIL import Image
import torch

# 加载模型和处理器
model_id = "starvector/starvector-8b-im2svg"
processor = AutoProcessor.from_pretrained(model_id)
model = VisionEncoderDecoderModel.from_pretrained(model_id)

device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

# 定义推理函数
def im2svg(image):
    inputs = processor(images=image, return_tensors="pt").to(device)
    outputs = model.generate(**inputs, max_new_tokens=1024)
    generated_svg = processor.batch_decode(outputs, skip_special_tokens=True)[0]
    return generated_svg

# 创建 Gradio 界面
demo = gr.Interface(
    fn=im2svg,
    inputs=gr.Image(type="pil"),
    outputs="text",
    title="StarVector 8B - Image to SVG",
    description="上传图像，将其转化为 SVG 矢量代码。",
    allow_flagging="never"
)

demo.launch()