第一章：模型加载与基础生成
1.1 基础模型加载
from diffusers import StableDiffusionPipeline
import torch
# 加载SD 1.5基础模型（FP32精度）
pipe = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float32
).to("cuda")
# 生成第一张图片
image = pipe("a cat wearing sunglasses").images[0]
image.save("basic_cat.png")
1.2 半精度优化加载
# 加载SDXL模型（FP16精度 + xformers加速）
pipe = StableDiffusionXLPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
torch_dtype=torch.float16,
use_xformers=True
).to("cuda")
# 生成高分辨率图片
image = pipe("cyberpunk city at night, 8k").images[0]
image.save("cyber_city.png")
1.3 低显存设备适配
# CPU卸载模式（显存<6GB适用）
pipe = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16
)
pipe.enable_model_cpu_offload() # 动态加载模型到显存
image = pipe("watercolor landscape").images[0]
image.save("low_vram_image.png")
第二章：Prompt工程优化
2.1 结构化Prompt模板
template = """
(cute corgi:1.3) wearing {glasses|hat|scarf}, # 主体
Studio Ghibli style, soft shading, # 风格
in a flower field at sunset, # 环境
4k resolution, bokeh effect # 画质
[blurry, low quality] # 负面提示
"""
image = pipe(template.format("sunglasses")).images[0]
image.save("styled_corgi.png")
2.2 动态权重控制
# 使用数值权重调整元素重要性
prompt = """
(a beautiful castle:1.5) on a cliff,
(medieval style:0.8) with (futuristic elements:0.6),
intricate details, 8k cinematic lighting
"""
image = pipe(prompt, guidance_scale=7).images[0]
image.save("hybrid_castle.png")
2.3 多语言Prompt融合
# 中英混合Prompt（需CLIP多语言支持）
prompt = """
A girl in 汉服 (hanfu:1.2) standing by 西湖 (West Lake),
水墨画风格 (ink wash painting style), ultra detailed
"""
image = pipe(prompt).images[0]
image.save("hanfu_girl.png")
第三章：高级控制技术
3.1 ControlNet姿势控制
from diffusers import ControlNetModel, StableDiffusionControlNetPipeline
from PIL import Image
# 加载OpenPose ControlNet
controlnet = ControlNetModel.from_pretrained(
"lllyasviel/sd-controlnet-openpose",
torch_dtype=torch.float16
)
pipe = StableDiffusionControlNetPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
controlnet=controlnet,
torch_dtype=torch.float16
).to("cuda")
# 输入姿势图
pose_image = Image.open("pose_ref.png")
image = pipe("dancing woman", image=pose_image).images[0]
image.save("controlled_dance.png")
3.2 LoRA风格叠加
from diffusers import StableDiffusionXLPipeline
import torch
# 加载基础模型
pipe = StableDiffusionXLPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
torch_dtype=torch.float16
).to("cuda")
# 加载动漫LoRA
pipe.load_lora_weights("lora/anime_style_xl.safetensors")
image = pipe("a warrior in armor").images[0]
image.save("anime_warrior.png")
3.3 多ControlNet联合控制
# 同时使用Canny边缘和深度图控制
controlnets = [
ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16),
ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-depth", torch_dtype=torch.float16)
]
pipe = StableDiffusionControlNetPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
controlnet=controlnets,
torch_dtype=torch.float16
).to("cuda")
# 输入多控制图
canny_img = Image.open("edge.png")
depth_img = Image.open("depth.png")
image = pipe("futuristic car", image=[canny_img, depth_img]).images[0]
image.save("multi_control_car.png")
第四章：性能优化
4.1 TensorRT加速
from diffusers import DiffusionPipeline
import torch_tensorrt
# 转换模型为TensorRT格式
pipe = DiffusionPipeline.from_pretrained(...)
trt_unet = torch_tensorrt.compile(
pipe.unet,
inputs=[torch.randn(1,4,64,64).to("cuda")],
enabled_precisions={torch.float16}
)
pipe.unet = trt_unet
# 加速生成
image = pipe("speed test image").images[0]
4.2 批处理生成
# 一次生成4张不同提示的图片
prompts = [
"a red rose",
"a blue rose",
"a golden rose",
"a black rose"
]
images = pipe(prompt=prompts, num_images_per_prompt=1).images
for idx, img in enumerate(images):
img.save(f"rose_{idx}.png")
4.3 缓存优化
from diffusers import StableDiffusionPipeline
import torch
# 启用KV缓存加速
pipe = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16,
enable_sequential_cpu_offload=True,
enable_kv_caching=True # 关键优化
).to("cuda")
# 第一次生成较慢（编译缓存）
image = pipe("warmup image").images[0]
# 后续生成加速30%
image = pipe("optimized image").images[0]
第五章：商业化生产
5.1 批量产品图生成
product_data = [
{"name": "sneakers", "color": "neon green"},
{"name": "backpack", "color": "matte black"},
{"name": "watch", "color": "rose gold"}
]
for product in product_data:
prompt = f"""
Professional product photo of {product['color']} {product['name']},
studio lighting, 8k resolution, product design award winner
"""
image = pipe(prompt).images[0]
image.save(f"{product['name']}_{product['color']}.png")
5.2 自动质量检测
from transformers import CLIPModel, CLIPProcessor
# 加载CLIP模型
clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
clip_processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
def quality_check(image, target_prompt):
inputs = clip_processor(text=target_prompt, images=image, return_tensors="pt")
outputs = clip_model(**inputs)
similarity = outputs.logits_per_image.item()
return similarity > 25 # 阈值根据实际情况调整
if quality_check(image, prompt):
image.save("approved.png")
else:
print("Quality check failed!")
5.3 多尺寸适配生成
resolutions = [(512,512), (768,768), (1024,1024)]
for w, h in resolutions:
pipe = StableDiffusionPipeline.from_pretrained(...)
image = pipe(
prompt,
width=w,
height=h,
target_size=(w,h)
).images[0]
image.save(f"output_{w}x{h}.png")
第六章：故障排查
6.1 显存监控
import nvidia_smi
nvidia_smi.nvmlInit()
handle = nvidia_smi.nvmlDeviceGetHandleByIndex(0)
def print_mem_usage():
info = nvidia_smi.nvmlDeviceGetMemoryInfo(handle)
print(f"Used VRAM: {info.used//1024**2} MB")
print_mem_usage() # 生成前后调用检测
6.2 异常处理
try:
image = pipe("problematic prompt").images[0]
except torch.cuda.OutOfMemoryError:
print("显存不足！尝试启用--medvram")
pipe.enable_model_cpu_offload()
image = pipe("problematic prompt").images[0]
6.3 采样调试
# 记录采样过程
pipe = StableDiffusionPipeline.from_pretrained(...)
pipe.set_progress_bar_config(leave=True) # 显示详细进度
# 生成并保存中间步骤
for i in range(pipe.scheduler.config.num_train_timesteps):
image = pipe(
prompt,
callback_on_step_end=lambda step, t, latents:
latents.save(f"step_{step}.pt")
)