AI绘画：从扩散原理到生产级出图的工程实践

import torch
from diffusers import StableDiffusionXLPipeline, DPMSolverMultistepScheduler

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16, variant="fp16",
).to("cuda")

# 换用 DPM-Solver++：20步即可达到 DDIM 50步的质量
pipe.scheduler = DPMSolverMultistepScheduler.from_config(
    pipe.scheduler.config, algorithm_type="dpmsolver++",
    use_karras_sigmas=True,          # Karras sigma 提升细节
)

prompt = ("commercial skincare poster background, product at lower third, "
          "large negative space on top, minimalist, soft natural light, "
          "champagne gold and light gray, studio quality, no text, 4k")
negative = "text, watermark, deformed, cluttered, lowres, extra fingers"

generator = torch.Generator("cuda").manual_seed(42)   # 种子=可复现

image = pipe(
    prompt=prompt, negative_prompt=negative,
    num_inference_steps=28, guidance_scale=7.5,
    width=1024, height=1536,
    generator=generator,
).images[0]

image.save("poster_bg.png")

from diffusers import StableDiffusionXLControlNetPipeline, ControlNetModel
from diffusers.utils import load_image
import cv2, numpy as np

controlnet = ControlNetModel.from_pretrained(
    "diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16)

pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    controlnet=controlnet, torch_dtype=torch.float16,
).to("cuda")

# 从参考图提取 Canny 边缘作为结构约束
ref = np.array(load_image("reference.jpg"))
edges = cv2.Canny(ref, 100, 200)
edges = np.stack([edges] * 3, axis=-1)   # 转三通道

out = pipe(
    prompt="modern product photography, same composition",
    image=edges,
    controlnet_conditioning_scale=0.8,   # 越大越贴合结构，过高会失真
    num_inference_steps=30,
).images[0]
out.save("controlled.png")

from diffusers import StableDiffusionXLPipeline
import torch

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16).to("cuda")

# 加载品牌风格 LoRA，权重决定风格强度
pipe.load_lora_weights("./brand_lora", adapter_name="brand")
pipe.set_adapters(["brand"], adapter_weights=[0.85])

img = pipe(prompt="<brand_style> a perfume bottle on marble",
           num_inference_steps=30, guidance_scale=7).images[0]
img.save("brand_style.png")

import json, hashlib, pathlib
from dataclasses import dataclass, asdict

@dataclass
class Job:
    prompt: str
    seed: int
    steps: int
    cfg: float
    lora_weight: float

def job_id(j: Job) -> str:
    return hashlib.md5(json.dumps(asdict(j), sort_keys=True).encode()).hexdigest()[:12]

def run_batch(jobs, out_dir="output"):
    pathlib.Path(out_dir).mkdir(exist_ok=True)
    for j in jobs:
        jid = job_id(j)
        if (pathlib.Path(out_dir) / f"{jid}.png").exists():
            continue                      # 幂等：已生成则跳过
        gen = torch.Generator("cuda").manual_seed(j.seed)
        img = pipe(prompt=j.prompt, num_inference_steps=j.steps,
                   guidance_scale=j.cfg, generator=gen).images[0]
        img.save(f"{out_dir}/{jid}.png")
        # 元数据落盘，保证可追溯
        with open(f"{out_dir}/{jid}.json", "w") as f:
            json.dump(asdict(j), f, ensure_ascii=False)