app.py

import gradio as gr
import torch
from diffusers import StableDiffusionPipeline

# 全域變數
pipe = None

def load_model_fn(model_id, device_choice):
    global pipe
    try:
        # 強制邏輯：如果是 CPU，一定要用 float32 才能跑
        if device_choice == "cpu":
            dtype = torch.float32
        else:
            dtype = torch.float16 if torch.cuda.is_available() else torch.float32
        
        yield f"⏳ 正在載入模型至 {device_choice.upper()}... (這可能需要幾分鐘)"
        
        pipe = StableDiffusionPipeline.from_pretrained(
            model_id, 
            torch_dtype=dtype,
            use_safetensors=True
        )
        pipe.to(device_choice)
        
        yield f"✅ 成功載入：{model_id} ({device_choice.upper()})"
    except Exception as e:
        yield f"❌ 載_入失敗：{str(e)}"

def generate_fn(prompt, steps, guidance):
    global pipe
    if pipe is None:
        return None, "⚠️ 請先點擊『載入模型』！"
    
    try:
        image = pipe(
            prompt, 
            num_inference_steps=int(steps), 
            guidance_scale=guidance
        ).images[0] # 確保回傳單張圖片
        return image, "✨ 生成成功！"
    except Exception as e:
        return None, f"❌ 錯誤：{str(e)}"

with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 🛠️ CPU 優化版 Stable Diffusion 控制台")
    
    with gr.Row():
        with gr.Column(scale=1):
            gr.Markdown("### 1. 模型設定")
            model_input = gr.Textbox(
                label="Hugging Face Model ID", 
                value="runwayml/stable-diffusion-v1-5"
            )
            # 這裡預設改為 "cpu"
            device_radio = gr.Radio(["cpu", "cuda"], value="cpu", label="執行設備 (預設 CPU)")
            load_btn = gr.Button("🔄 載入模型", variant="secondary")
            load_status = gr.Markdown("系統狀態：等待指令")
            
            gr.Markdown("### 2. 繪圖參數")
            prompt = gr.Textbox(label="提示詞 (Prompt)", lines=3, placeholder="An astronaut riding a horse")
            steps = gr.Slider(1, 50, value=20, step=1, label="步數 (CPU 建議 15-20)")
            guidance = gr.Slider(1, 20, value=7.5, step=0.5, label="提示詞強度")
            gen_btn = gr.Button("🚀 開始生成", variant="primary")

        with gr.Column(scale=2):
            output_img = gr.Image(label="生成結果")
            status_msg = gr.Textbox(label="執行訊息", interactive=False)

    load_btn.click(fn=load_model_fn, inputs=[model_input, device_radio], outputs=load_status)
    gen_btn.click(fn=generate_fn, inputs=[prompt, steps, guidance], outputs=[output_img, status_msg])

demo.launch()