Upload app.py with huggingface_hub

app.py

@@ -0,0 +1,231 @@
+import os
+import torch
+import gradio as gr
+from diffusers import StableDiffusionXLPipeline, AutoencoderKL
+from diffusers.schedulers import EulerAncestralDiscreteScheduler
+from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
+from safetensors.torch import load_file
+import gc
+
+# ─── Config ──────────────────────────────────────────────────────────
+MODEL_ID = "NoobAI-XL/NoobAI-XL-1.1-SDNQ-UINT4"  # HF repo
+DEVICE = "cpu"
+DTYPE = torch.float32  # CPU 用 fp32 更稳
+DEFAULT_W, DEFAULT_H = 640, 960
+DEFAULT_STEPS = 20
+DEFAULT_GUIDANCE = 5.0
+
+# ─── Prompt enhancement (Chinese → English + quality tags) ───────────
+QUALITY_TAGS = (
+    "masterpiece, best quality, ultra high res, "
+    "dynamic composition, dramatic lighting, "
+    "detailed background, manga style, anime style, "
+    "sharp focus, vivid colors, professional illustration"
+)
+
+NEGATIVE_PROMPT = (
+    "low quality, worst quality, bad anatomy, bad composition, "
+    "blurry, pixelated, watermark, text, signature, logo, "
+    "ugly, deformed, noise, artifacts, lowres, error, "
+    "missing fingers, extra digits, fewer digits, cropped, "
+    "jpeg artifacts, out of frame, mutation, mutated, "
+    "extra limbs, malformed limbs, long neck, disfigured"
+)
+
+STYLE_PROMPTS = {
+    "Manga / Anime": "manga style, anime style, screentone, ink lines, dynamic panel composition",
+    "Western Comics": "western comic style, bold outlines, halftone dots, vibrant colors, dynamic action",
+    "Manhwa": "manhwa style, webtoon style, clean lines, soft shading, vertical composition",
+    "Sketch / Lineart": "sketch style, lineart, black and white, pencil drawing, detailed strokes",
+}
+
+def enhance_prompt(user_prompt: str, style: str) -> str:
+    """中文场景 → 英文提示词 + 质量标签 + 风格标签"""
+    # 简单的中文关键词映射（可扩展为更完整的翻译/扩写）
+    cn_en_map = {
+        "恐龙": "dinosaur",
+        "霸王龙": "T-Rex",
+        "暴雨": "heavy rain",
+        "史前森林": "prehistoric forest",
+        "仰头": "looking up",
+        "怒吼": "roaring",
+        "闪电": "lightning",
+        "照亮": "illuminating",
+        "天空": "sky",
+        "低角度": "low angle",
+        "镜头": "camera angle",
+        "强烈": "intense",
+        "漫画": "manga",
+        "构图": "composition",
+        "动态": "dynamic",
+        "线条": "lines",
+        "丰富": "rich",
+        "背景": "background",
+        "细节": "details",
+        "愤怒": "angry",
+        "站在": "standing in",
+        "里": "in",
+        "的": "",
+        "一只": "a",
+        "在": "in",
+        "中": "in",
+        "和": "and",
+        "与": "and",
+    }
+    # 粗略翻译：把常见中文词替换成英文
+    translated = user_prompt
+    for cn, en in cn_en_map.items():
+        translated = translated.replace(cn, en)
+    # 如果全是中文，做简单拼接；如果已经有英文，保留
+    if any('\u4e00' <= c <= '\u9fff' for c in translated):
+        # 还含有中文，简单处理
+        prompt_en = translated
+    else:
+        prompt_en = translated
+    style_tag = STYLE_PROMPTS.get(style, STYLE_PROMPTS["Manga / Anime"])
+    return f"{prompt_en}, {style_tag}, {QUALITY_TAGS}"
+
+# ─── Load pipeline (lazy, cached) ────────────────────────────────────
+_pipe = None
+
+def get_pipeline():
+    global _pipe
+    if _pipe is not None:
+        return _pipe
+
+    print("Loading NoobAI-XL SDNQ pipeline on CPU...")
+
+    # 1) VAE (单独加载，启用 slicing)
+    vae = AutoencoderKL.from_pretrained(
+        MODEL_ID,
+        subfolder="vae",
+        torch_dtype=DTYPE,
+        low_cpu_mem_usage=True,
+    )
+    vae.enable_slicing()
+    vae.enable_tiling()
+
+    # 2) Text encoders & tokenizers (SDXL 双编码器)
+    tokenizer_1 = CLIPTokenizer.from_pretrained(MODEL_ID, subfolder="tokenizer")
+    tokenizer_2 = CLIPTokenizer.from_pretrained(MODEL_ID, subfolder="tokenizer_2")
+    text_encoder_1 = CLIPTextModel.from_pretrained(
+        MODEL_ID, subfolder="text_encoder", torch_dtype=DTYPE, low_cpu_mem_usage=True
+    )
+    text_encoder_2 = CLIPTextModelWithProjection.from_pretrained(
+        MODEL_ID, subfolder="text_encoder_2", torch_dtype=DTYPE, low_cpu_mem_usage=True
+    )
+
+    # 3) Pipeline
+    pipe = StableDiffusionXLPipeline.from_pretrained(
+        MODEL_ID,
+        vae=vae,
+        text_encoder=text_encoder_1,
+        text_encoder_2=text_encoder_2,
+        tokenizer=tokenizer_1,
+        tokenizer_2=tokenizer_2,
+        scheduler=EulerAncestralDiscreteScheduler.from_pretrained(MODEL_ID, subfolder="scheduler"),
+        torch_dtype=DTYPE,
+        low_cpu_mem_usage=True,
+        variant="fp16" if DTYPE == torch.float16 else None,
+    )
+
+    # CPU 优化
+    pipe.enable_model_cpu_offload()           # 逐层卸载到 CPU
+    pipe.enable_vae_slicing()
+    pipe.enable_vae_tiling()
+    pipe.enable_attention_slicing("auto")     # 注意力切片
+    # 关闭 xformers / flash attention（CPU 不支持）
+    try:
+        pipe.disable_xformers_memory_efficient_attention()
+    except Exception:
+        pass
+
+    pipe.to(DEVICE)
+    _pipe = pipe
+    print("Pipeline ready.")
+    return pipe
+
+# ─── Generation ──────────────────────────────────────────────────────
+def generate(prompt, style, width, height, steps, guidance, progress=gr.Progress()):
+    if not prompt.strip():
+        return None, "请输入场景描述"
+
+    progress(0.1, desc="增强提示词...")
+    full_prompt = enhance_prompt(prompt, style)
+
+    progress(0.2, desc="加载模型...")
+    pipe = get_pipeline()
+
+    progress(0.4, desc="生成中（CPU 慢，请耐心等待）...")
+    with torch.inference_mode():
+        image = pipe(
+            prompt=full_prompt,
+            negative_prompt=NEGATIVE_PROMPT,
+            width=width,
+            height=height,
+            num_inference_steps=steps,
+            guidance_scale=guidance,
+            num_images_per_prompt=1,
+        ).images[0]
+
+    progress(0.9, desc="保存 PNG...")
+    # 保存到临时文件供 Gradio 返回
+    out_path = "/tmp/comic_scene.png"
+    image.save(out_path, "PNG")
+    progress(1.0, desc="完成")
+    gc.collect()
+    torch.cuda.empty_cache() if torch.cuda.is_available() else None
+    return out_path, f"✅ 生成完成\nPrompt: {full_prompt[:200]}..."
+
+# ─── Gradio UI ───────────────────────────────────────────────────────
+css = """
+#title { text-align: center; font-size: 1.5rem; font-weight: 700; margin-bottom: 0.5rem; }
+#subtitle { text-align: center; color: #666; margin-bottom: 1rem; }
+.generate-btn { background: linear-gradient(90deg, #667eea 0%, #764ba2 100%) !important; }
+"""
+
+with gr.Blocks(css=css, title="Comic Scene Generator") as demo:
+    gr.HTML('<div id="title">🎨 Comic Scene Generator</div>')
+    gr.HTML('<div id="subtitle">输入一句中文场景 → 自动扩写 → NoobAI-XL 生成漫画场景图</div>')
+
+    with gr.Row():
+        with gr.Column(scale=1):
+            prompt = gr.Textbox(
+                label="Scene",
+                placeholder="例：一只愤怒的霸王龙站在暴雨中的史前森林里仰头怒吼，闪电照亮天空，低角度镜头，强烈的漫画构图，动态线条，丰富背景细节",
+                lines=4,
+            )
+            style = gr.Dropdown(
+                label="Style",
+                choices=list(STYLE_PROMPTS.keys()),
+                value="Manga / Anime",
+            )
+            with gr.Row():
+                width = gr.Number(label="Width", value=DEFAULT_W, precision=0, minimum=320, maximum=1024, step=64)
+                height = gr.Number(label="Height", value=DEFAULT_H, precision=0, minimum=320, maximum=1536, step=64)
+            steps = gr.Slider(label="Steps", minimum=10, maximum=50, value=DEFAULT_STEPS, step=1)
+            guidance = gr.Slider(label="Guidance Scale", minimum=1.0, maximum=15.0, value=DEFAULT_GUIDANCE, step=0.5)
+            btn = gr.Button("Generate", elem_classes="generate-btn", variant="primary")
+
+        with gr.Column(scale=1):
+            output_img = gr.Image(label="生成的漫画场景", type="filepath", interactive=False)
+            status = gr.Textbox(label="Status", interactive=False)
+
+    btn.click(
+        generate,
+        inputs=[prompt, style, width, height, steps, guidance],
+        outputs=[output_img, status],
+    )
+
+    # 示例
+    gr.Examples(
+        examples=[
+            ["一只愤怒的霸王龙站在暴雨中的史前森林里仰头怒吼，闪电照亮天空，低角度镜头，强烈的漫画构图，动态线条，丰富背景细节", "Manga / Anime", 640, 960, 20, 5.0],
+            ["赛博朋克城市夜晚，霓虹灯招牌倒映在湿漉漉的街道上，孤独的少女撑着透明伞，电影感构图", "Manga / Anime", 640, 960, 20, 5.0],
+            ["古风仙侠，白衣剑仙御剑飞行穿越云海，日出金光洒在云层上，宏大壮观，工笔画风", "Manga / Anime", 640, 960, 25, 5.5],
+        ],
+        inputs=[prompt, style, width, height, steps, guidance],
+    )
+
+if __name__ == "__main__":
+    demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True)