Create app.py

app.py

@@ -0,0 +1,232 @@
+"""
+Z-Image 图像生成演示
+简化版 UI，移除了提示词优化逻辑，合并了分辨率选择。
+"""
+
+import spaces
+import random
+import re
+import torch
+import gradio as gr
+from diffusers import ZImagePipeline
+
+# ==================== 配置信息 ====================
+MODEL_PATH = "Tongyi-MAI/Z-Image"
+
+# ==================== 合并后的分辨率列表 ====================
+ALL_RESOLUTIONS = [
+    # 720 级别
+    "720x720 ( 1:1 )",
+    "896x512 ( 16:9 )",
+    "512x896 ( 9:16 )",
+    "832x544 ( 3:2 )",
+    "544x832 ( 2:3 )",
+    "800x576 ( 4:3 )",
+    "576x800 ( 3:4 )",
+    # 1024 级别
+    "1024x1024 ( 1:1 )",
+    "1152x896 ( 9:7 )",
+    "896x1152 ( 7:9 )",
+    "1152x864 ( 4:3 )",
+    "864x1152 ( 3:4 )",
+    "1248x832 ( 3:2 )",
+    "832x1248 ( 2:3 )",
+    "1280x720 ( 16:9 )",
+    "720x1280 ( 9:16 )",
+    "1344x576 ( 21:9 )",
+    "576x1344 ( 9:21 )",
+    # 1280 级别
+    "1280x1280 ( 1:1 )",
+    "1440x1120 ( 9:7 )",
+    "1120x1440 ( 7:9 )",
+    "1472x1104 ( 4:3 )",
+    "1104x1472 ( 3:4 )",
+    "1536x1024 ( 3:2 )",
+    "1024x1536 ( 2:3 )",
+    "1536x864 ( 16:9 )",
+    "864x1536 ( 9:16 )",
+    "1680x720 ( 21:9 )",
+    "720x1680 ( 9:21 )",
+]
+
+EXAMPLE_PROMPTS = [
+    ["一位男士和他的贵宾犬穿着配套的服装参加狗狗秀，室内灯光，背景中有观众。"],
+    ["极具氛围感的暗调人像，一位优雅的中国美女在黑暗的房间里。一束强光通过遮光板，在她的脸上投射出一个清晰的闪电形状的光影。"],
+    ["Young Chinese woman in red Hanfu, intricate embroidery, golden phoenix headdress, soft-lit outdoor night background."],
+    ["A serene mountain landscape at sunset with golden light reflecting off a calm lake."],
+    ["A futuristic cityscape with flying cars and neon holographic advertisements, cyberpunk style."],
+]
+
+# ==================== 辅助函数 ====================
+def get_resolution(resolution: str) -> tuple[int, int]:
+    """解析分辨率字符串为宽度和高度。"""
+    match = re.search(r"(\d+)\s*[×x]\s*(\d+)", resolution)
+    if match:
+        return int(match.group(1)), int(match.group(2))
+    return 1024, 1024
+
+# ==================== 模型加载 ====================
+print(f"正在从 {MODEL_PATH} 加载 Z-Image 流水线...")
+pipe = ZImagePipeline.from_pretrained(
+    MODEL_PATH,
+    torch_dtype=torch.bfloat16,
+    low_cpu_mem_usage=False,
+)
+pipe.to("cuda")
+print("流水线加载成功！")
+
+# ==================== 生成核心逻辑 ====================
+@spaces.GPU
+def generate(
+    prompt: str,
+    negative_prompt: str = "",
+    resolution: str = "1024x1024 ( 1:1 )",
+    seed: int = 42,
+    num_inference_steps: int = 30,
+    guidance_scale: float = 4.0,
+    cfg_normalization: bool = False,
+    random_seed: bool = True,
+    gallery_images: list = None,
+    progress=gr.Progress(track_tqdm=True),
+):
+    """
+    使用 Z-Image 扩散模型生成图像。
+    仅保留核心生成逻辑。
+    """
+    
+    if not prompt.strip():
+        raise gr.Error("请输入提示词。")
+    
+    # 处理种子
+    if random_seed:
+        new_seed = random.randint(1, 1000000)
+    else:
+        new_seed = seed if seed != -1 else random.randint(1, 1000000)
+    
+    # 解析分辨率
+    width, height = get_resolution(resolution)
+    
+    # 设置生成器
+    generator = torch.Generator("cuda").manual_seed(new_seed)
+    
+    # 执行生成
+    image = pipe(
+        prompt=prompt,
+        negative_prompt=negative_prompt if negative_prompt.strip() else None,
+        height=height,
+        width=width,
+        cfg_normalization=cfg_normalization,
+        num_inference_steps=num_inference_steps,
+        guidance_scale=guidance_scale,
+        generator=generator,
+    ).images[0]
+    
+    # 更新画廊
+    if gallery_images is None:
+        gallery_images = []
+    gallery_images = [image] + gallery_images
+    
+    return gallery_images, str(new_seed), int(new_seed)
+
+# ==================== Gradio 界面设计 ====================
+with gr.Blocks(title="Z-Image 核心生成器") as demo:
+    gr.Markdown(
+        """<div align="center">
+        <h1>Z-Image 核心生成器</h1>
+        <p>基于单流扩散 Transformer 的高效图像生成模型</p>
+        </div>"""
+    )
+
+    with gr.Row():
+        with gr.Column(scale=1):
+            prompt_input = gr.Textbox(
+                label="提示词 (Prompt)",
+                lines=4,
+                placeholder="在此输入你想生成的画面描述（支持中英文）..."
+            )
+            negative_prompt_input = gr.Textbox(
+                label="反向提示词 (可选)",
+                lines=2,
+                placeholder="输入你不想在图像中出现的内容..."
+            )
+
+            # 分辨率改为单一完整的下拉列表
+            resolution = gr.Dropdown(
+                label="分辨率选择 (宽 x 高)",
+                choices=ALL_RESOLUTIONS,
+                value="1024x1024 ( 1:1 )"
+            )
+
+            with gr.Row():
+                seed = gr.Number(label="种子", value=42, precision=0)
+                random_seed = gr.Checkbox(label="使用随机种子", value=True)
+
+            with gr.Row():
+                num_inference_steps = gr.Slider(
+                    label="推理步数 (Steps)",
+                    minimum=10,
+                    maximum=100,
+                    value=30,
+                    step=1
+                )
+                guidance_scale = gr.Slider(
+                    label="引导比例 (CFG Scale)",
+                    minimum=1.0,
+                    maximum=20.0,
+                    value=4.0,
+                    step=0.5
+                )
+
+            cfg_normalization = gr.Checkbox(
+                label="启用 CFG 归一化",
+                value=False
+            )
+
+            generate_btn = gr.Button("开始生成", variant="primary")
+
+            # 示例提示词
+            gr.Markdown("### 📝 示例")
+            gr.Examples(
+                examples=EXAMPLE_PROMPTS,
+                inputs=prompt_input,
+                label=None
+            )
+
+        with gr.Column(scale=1):
+            output_gallery = gr.Gallery(
+                label="生成结果",
+                columns=1,
+                rows=1,
+                height=512,
+                object_fit="contain",
+                format="png",
+                interactive=False,
+            )
+            used_seed = gr.Textbox(label="本次使用的种子", interactive=False)
+
+    # 绑定生成按钮事件
+    generate_btn.click(
+        fn=generate,
+        inputs=[
+            prompt_input,
+            negative_prompt_input,
+            resolution,
+            seed,
+            num_inference_steps,
+            guidance_scale,
+            cfg_normalization,
+            random_seed,
+            output_gallery,
+        ],
+        outputs=[output_gallery, used_seed, seed],
+        api_name="generate",
+    )
+
+# ==================== 启动 ====================
+css = ".fillable{max-width: 1230px !important}"
+if __name__ == "__main__":
+    demo.launch(
+        server_name="0.0.0.0",
+        server_port=7860,
+        css=css
+    )