Python AI 绘画
Calling image APIs, running local diffusion models, and handling the async/polling pattern.
Key PointsImage APIs are slow (5-30s) — treat them as async jobs with polling, not as a synchronous request in your web handler.
Call an Image API
APIs return either a URL or base64. Base64 can be written straight to disk; URLs expire and may be hotlink-protected.
Python
import base64
from pathlib import Path
from openai import OpenAI
client = OpenAI(timeout=120)
resp = client.images.generate(
model="gpt-image-1",
prompt="a minimal isometric illustration of a database server, flat colors",
size="1024x1024",
n=1,
)
b64 = resp.data[0].b64_json
Path("out.png").write_bytes(base64.b64decode(b64))
print("saved", Path("out.png").stat().st_size, "bytes")Output
saved 1048576 bytesPost-processing with Pillow
Common post-processing — resize, crop, watermark, convert to WebP — is one Pillow chain; OpenCV is overkill.
Python
from PIL import Image
img = Image.open("out.png").convert("RGB")
img.thumbnail((512, 512)) # 等比缩放，不放大
img.save("out.webp", "WEBP", quality=82)
print(img.size)Output
(512, 512)Local Diffusion
Running Stable Diffusion locally needs GPU VRAM (6GB+). diffusers is the usual wrapper; first run downloads gigabytes of weights.
Python
from diffusers import StableDiffusionPipeline
import torch
pipe = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16,
).to("cuda")
image = pipe("a cat sitting on a server rack, studio light", num_inference_steps=25).images[0]
image.save("cat.png")Never generate images synchronously inside a web request — 10-30s per call saturates workers. Queue it, return a job id, poll or callback.
调用图片接口
接口返回的一般是图片 URL 或 base64。base64 直接落盘更快，URL 则要考虑有效期和防盗链。
Python
import base64
from pathlib import Path
from openai import OpenAI
client = OpenAI(timeout=120)
resp = client.images.generate(
model="gpt-image-1",
prompt="a minimal isometric illustration of a database server, flat colors",
size="1024x1024",
n=1,
)
b64 = resp.data[0].b64_json
Path("out.png").write_bytes(base64.b64decode(b64))
print("saved", Path("out.png").stat().st_size, "bytes")Output
saved 1048576 bytes用 Pillow 后处理
生成图常见需求是压缩、裁剪、加水印、转 WebP。Pillow 一条链就能做完，不必引入 OpenCV。
Python
from PIL import Image
img = Image.open("out.png").convert("RGB")
img.thumbnail((512, 512)) # 等比缩放，不放大
img.save("out.webp", "WEBP", quality=82)
print(img.size)Output
(512, 512)本地扩散模型
本地跑 Stable Diffusion 需要 GPU 显存（一般 6GB 起）。diffusers 是最常用的封装，首次运行会下载数 GB 权重。
Python
from diffusers import StableDiffusionPipeline
import torch
pipe = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16,
).to("cuda")
image = pipe("a cat sitting on a server rack, studio light", num_inference_steps=25).images[0]
image.save("cat.png")不要在 Web 请求线程里同步生成图片：一次 10~30 秒会占满 worker。正确做法是入队 + 任务 ID + 轮询/回调。