import gradio as gr
from transformers import (
    SiglipImageProcessor,
    RobertaTokenizerFast,
    VisionEncoderDecoderModel
)
from PIL import Image
import torch

# 模型 ID
model_id = "starvector/starvector-8b-im2svg"

# 分别加载 image processor 和 tokenizer
image_processor = SiglipImageProcessor.from_pretrained(model_id)
tokenizer = RobertaTokenizerFast.from_pretrained(model_id)
model = VisionEncoderDecoderModel.from_pretrained(model_id)

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

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

# Gradio UI
demo = gr.Interface(
    fn=im2svg,
    inputs=gr.Image(type="pil"),
    outputs="text",
    title="🖼️ StarVector: Image → SVG",
    description="上传图像，我将它转化为矢量图（SVG 代码）。适用于简笔画、图标、草图。",
)

demo.launch()