app.py

import gradio as gr
from model_pipelines import load_pipelines, generate_all
from grace_eval import compute_sample_scores, plot_radar

pipes = load_pipelines()

def compare(prompt):
    imgs = generate_all(pipes, prompt)
    return imgs["sd_v1_5"], imgs["openjourney_v4"], imgs["ldm_256"]

def show_leaderboard(prompt):
    scores = compute_sample_scores(None, prompt)
    plot_radar(scores)
    return "radar.png"

with gr.Blocks() as demo:
    gr.Markdown("# 图像生成模型对比实验")

    with gr.Tab("Arena"):
        prompt = gr.Textbox(label="请输入生成文本")
        btn = gr.Button("生成图像")
        out1 = gr.Image(label="StableDiffusion v1.5")
        out2 = gr.Image(label="Openjourney v4")
        out3 = gr.Image(label="LDM 256")
        btn.click(compare, inputs=prompt, outputs=[out1, out2, out3])

    with gr.Tab("Leaderboard"):
        pm = gr.Textbox(label="统一 prompt 用于评价")
        lb_btn = gr.Button("显示 GRACE 雷达图")
        lb_img = gr.Image()
        lb_btn.click(show_leaderboard, inputs=pm, outputs=lb_img)

    with gr.Tab("Report"):
        markdown = open("report.md", "r", encoding="utf-8").read()
        gr.Markdown(markdown)

demo.launch()