Python 实现 AI 图像生成：调用 Stable Diffusion API 完整教程

优质文章学习记录

— 8 min read
Python 实现 AI 图像生成：调用 Stable Diffusion API 完整教程

从零开始学习使用 Python 调用 Stable Diffusion API 生成图像，涵盖本地部署、API 调用、ControlNet、图生图等进阶技巧。

1. 技术架构

Python 客户端

Stable Diffusion API

本地部署
SD WebUI / ComfyUI

云端 API
Replicate / Stability AI

Stable Diffusion 模型

文生图
txt2img

图生图
img2img

局部重绘
inpainting

超分辨率
upscale

输出图像

后处理管道

存储
本地/OSS

2. 图像生成方式对比

50%25%15%10%各生成方式使用占比统计文生图 (txt2img)图生图 (img2img)局部重绘 (inpainting)超分辨率 (upscale)

3. 环境准备

3.1 本地部署 Stable Diffusion WebUI

# 克隆 Stable Diffusion WebUIgit clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git cd stable-diffusion-webui # 启动（开启 API 模式） ./webui.sh --api--listen# Windows 用户 webui.bat --api--listen

3.2 安装依赖

pip install requests Pillow io base64

4. 核心代码实现

4.1 SD API 客户端封装

# sd_client.pyimport requests import base64 import io import json import time from pathlib import Path from PIL import Image from dataclasses import dataclass, field from typing import Optional @dataclassclassGenerationConfig:"""图像生成配置""" prompt:str="" negative_prompt:str="low quality, blurry, deformed" width:int=512 height:int=512 steps:int=30 cfg_scale:float=7.0 sampler_name:str="DPM++ 2M Karras" seed:int=-1# -1 表示随机 batch_size:int=1 n_iter:int=1# 迭代次数 model: Optional[str]=NoneclassStableDiffusionClient:"""Stable Diffusion API 客户端"""def__init__(self, base_url:str="http://127.0.0.1:7860"): self.base_url = base_url self.api_url =f"{base_url}/sdapi/v1"def_save_base64_image(self, b64_str:str, output_path:str)->str:"""将 base64 图片保存到文件""" img_data = base64.b64decode(b64_str) img = Image.open(io.BytesIO(img_data)) img.save(output_path)return output_path # ---- 文生图 ----deftxt2img(self, config: GenerationConfig, output_dir:str="./output")->list[str]:"""文生图：从文本描述生成图像""" payload ={"prompt": config.prompt,"negative_prompt": config.negative_prompt,"width": config.width,"height": config.height,"steps": config.steps,"cfg_scale": config.cfg_scale,"sampler_name": config.sampler_name,"seed": config.seed,"batch_size": config.batch_size,"n_iter": config.n_iter,}if config.model: self._switch_model(config.model) response = requests.post(f"{self.api_url}/txt2img", json=payload) response.raise_for_status() data = response.json() Path(output_dir).mkdir(exist_ok=True) saved_paths =[]for i, img_b64 inenumerate(data["images"]): path =f"{output_dir}/txt2img_{int(time.time())}_{i}.png" self._save_base64_image(img_b64, path) saved_paths.append(path)print(f"已保存: {path}")return saved_paths # ---- 图生图 ----defimg2img(self, init_image_path:str, prompt:str, denoising_strength:float=0.75, config: GenerationConfig =None, output_dir:str="./output")->list[str]:"""图生图：基于参考图 + 提示词生成新图""" config = config or GenerationConfig()# 读取初始图片并转 base64withopen(init_image_path,"rb")as f: init_images =[base64.b64encode(f.read()).decode()] payload ={"init_images": init_images,"prompt": prompt,"negative_prompt": config.negative_prompt,"width": config.width,"height": config.height,"steps": config.steps,"cfg_scale": config.cfg_scale,"sampler_name": config.sampler_name,"denoising_strength": denoising_strength,"seed": config.seed,} response = requests.post(f"{self.api_url}/img2img", json=payload) response.raise_for_status() data = response.json() Path(output_dir).mkdir(exist_ok=True) saved_paths =[]for i, img_b64 inenumerate(data["images"]): path =f"{output_dir}/img2img_{int(time.time())}_{i}.png" self._save_base64_image(img_b64, path) saved_paths.append(path)print(f"已保存: {path}")return saved_paths # ---- 局部重绘 ----definpaint(self, init_image_path:str, mask_image_path:str, prompt:str, denoising_strength:float=0.85, output_dir:str="./output")->list[str]:"""局部重绘：只修改 mask 区域"""withopen(init_image_path,"rb")as f: init_images =[base64.b64encode(f.read()).decode()]withopen(mask_image_path,"rb")as f: mask = base64.b64encode(f.read()).decode() payload ={"init_images": init_images,"mask": mask,"prompt": prompt,"negative_prompt":"low quality, blurry","denoising_strength": denoising_strength,"inpainting_fill":1,# 0=fill, 1=original, 2=latent noise"inpaint_full_res":True,"steps":30,"cfg_scale":7.0,"sampler_name":"DPM++ 2M Karras","width":512,"height":512,} response = requests.post(f"{self.api_url}/img2img", json=payload) response.raise_for_status() data = response.json() Path(output_dir).mkdir(exist_ok=True) saved_paths =[]for i, img_b64 inenumerate(data["images"]): path =f"{output_dir}/inpaint_{int(time.time())}_{i}.png" self._save_base64_image(img_b64, path) saved_paths.append(path)return saved_paths # ---- 超分辨率 ----defupscale(self, image_path:str, scale:int=2, output_dir:str="./output")->str:"""使用 ESRGAN 进行超分辨率放大"""withopen(image_path,"rb")as f: img_b64 = base64.b64encode(f.read()).decode() payload ={"image": img_b64,"upscaler_1":"R-ESRGAN 4x+","upscaling_resize": scale,} response = requests.post(f"{self.api_url}/extra-single-image", json=payload) response.raise_for_status() data = response.json() Path(output_dir).mkdir(exist_ok=True) path =f"{output_dir}/upscaled_{int(time.time())}.png" self._save_base64_image(data["image"], path)print(f"超分辨率完成: {path}")return path # ---- 模型管理 ----def_switch_model(self, model_name:str):"""切换模型""" response = requests.post(f"{self.api_url}/options", json={"sd_model_checkpoint": model_name},) response.raise_for_status() time.sleep(3)# 等待模型加载deflist_models(self)->list[str]:"""列出可用模型""" response = requests.get(f"{self.api_url}/sd-models")return[m["title"]for m in response.json()]deflist_samplers(self)->list[str]:"""列出可用采样器""" response = requests.get(f"{self.api_url}/samplers")return[s["name"]for s in response.json()]

4.2 批量生成示例

# batch_generate.pyfrom sd_client import StableDiffusionClient, GenerationConfig defbatch_generate_portraits():"""批量生成人物肖像""" sd = StableDiffusionClient()# 查看可用模型和采样器print("可用模型:", sd.list_models()[:5])print("可用采样器:", sd.list_samplers())# 风格列表 styles =["cyberpunk neon city","watercolor painting","oil painting renaissance","anime style","photorealistic 8k",] base_prompt =("portrait of a young woman, detailed face, beautiful eyes, ""dramatic lighting, masterpiece, best quality")for style in styles: config = GenerationConfig( prompt=f"{base_prompt}, {style}", negative_prompt="lowres, bad anatomy, bad hands, text, error", width=512, height=768, steps=30, cfg_scale=7.5,) paths = sd.txt2img(config, output_dir=f"./output/{style.replace(' ','_')}")print(f"风格 [{style}] -> {paths}")if __name__ =="__main__": batch_generate_portraits()

4.3 调用 Stability AI 云端 API

# stability_cloud.pyimport requests import base64 from pathlib import Path from PIL import Image from io import BytesIO classStabilityAIClient:"""Stability AI 官方云端 API"""def__init__(self, api_key:str): self.api_key = api_key self.base_url ="https://api.stability.ai/v2beta"defgenerate(self, prompt:str, aspect_ratio:str="1:1", style:str="photographic", output_path:str="output.png")->str:"""调用 Stable Diffusion 3 生成图像""" response = requests.post(f"{self.base_url}/stable-image/generate/sd3", headers={"Authorization":f"Bearer {self.api_key}","Accept":"image/*",}, files={"none":""}, data={"prompt": prompt,"aspect_ratio": aspect_ratio,"style_preset": style,"output_format":"png",},)if response.status_code !=200:raise Exception(f"API 错误: {response.status_code} - {response.text}")withopen(output_path,"wb")as f: f.write(response.content)print(f"已生成: {output_path}")return output_path # 使用示例if __name__ =="__main__": client = StabilityAIClient(api_key="sk-your-api-key") client.generate( prompt="A majestic dragon flying over a neon-lit cyberpunk city at night, ""highly detailed, cinematic lighting, 8k", aspect_ratio="16:9", style="cinematic", output_path="dragon_city.png",)

4.4 图像后处理管道

# postprocess.pyfrom PIL import Image, ImageEnhance, ImageFilter from pathlib import Path classImagePostProcessor:"""图像后处理：调整色彩、锐化、添加水印"""@staticmethoddefenhance(image_path:str, brightness:float=1.1, contrast:float=1.15, sharpness:float=1.3, output_path:str=None)->str:"""综合增强""" img = Image.open(image_path) img = ImageEnhance.Brightness(img).enhance(brightness) img = ImageEnhance.Contrast(img).enhance(contrast) img = ImageEnhance.Sharpness(img).enhance(sharpness) output_path = output_path or image_path.replace(".","_enhanced.") img.save(output_path, quality=95)return output_path @staticmethoddefadd_watermark(image_path:str, text:str="AI Generated", output_path:str=None)->str:"""添加水印"""from PIL import ImageDraw, ImageFont img = Image.open(image_path).convert("RGBA") overlay = Image.new("RGBA", img.size,(0,0,0,0)) draw = ImageDraw.Draw(overlay)# 半透明白色文字 draw.text((img.width -200, img.height -40), text, fill=(255,255,255,128),) img = Image.alpha_composite(img, overlay).convert("RGB") output_path = output_path or image_path.replace(".","_wm.") img.save(output_path, quality=95)return output_path @staticmethoddefcreate_grid(image_paths:list[str], cols:int=3, output_path:str="grid.png")->str:"""将多张图片拼成网格""" images =[Image.open(p)for p in image_paths] w, h = images[0].size rows =(len(images)+ cols -1)// cols grid = Image.new("RGB",(w * cols, h * rows),"white")for i, img inenumerate(images): row, col =divmod(i, cols) grid.paste(img,(col * w, row * h)) grid.save(output_path, quality=95)print(f"网格图已保存: {output_path}")return output_path

5. Prompt 工程技巧

Prompt 结构

主体描述

风格关键词

质量修饰词

负面提示词

高质量 Prompt 模板

PROMPT_TEMPLATES ={"人物肖像":("{subject}, {style}, detailed face, expressive eyes, ""dramatic lighting, masterpiece, best quality, ultra detailed"),"风景":("{scene}, {mood}, volumetric lighting, god rays, ""landscape photography, 8k uhd, cinematic composition"),"产品设计":("{product}, minimalist design, studio lighting, ""white background, product photography, professional, 4k"),"动漫":("{character}, anime style, vibrant colors, ""detailed illustration, cel shading, masterpiece"),} NEGATIVE_PROMPTS ={"通用":"lowres, bad anatomy, bad hands, text, error, missing fingers, ""extra digit, cropped, worst quality, low quality, blurry","写实":"illustration, painting, drawing, art, sketch, anime, cartoon, ""CG, render, 3D, watermark, text, font, signature","动漫":"photo, realistic, 3d, western, ugly, duplicate, morbid, ""deformed, bad anatomy, blurry",}

6. 关键参数影响

35%20%15%15%10%5%不同参数对生成质量的影响权重Prompt 质量采样步数 (steps)CFG Scale采样器选择模型选择分辨率

参数推荐值说明
steps25-35步数越多细节越好，但边际递减且更慢
cfg_scale7-12越高越遵循 prompt，过高会过饱和
samplerDPM++ 2M Karras兼顾速度与质量
denoising_strength0.5-0.8图生图降噪强度，越高变化越大
seed-1随机种子，固定可复现

7. 完整使用流程

# complete_demo.pyfrom sd_client import StableDiffusionClient, GenerationConfig from stability_cloud import StabilityAIClient from postprocess import ImagePostProcessor defmain():# ===== 方式一：本地 SD WebUI ===== sd = StableDiffusionClient("http://127.0.0.1:7860")# 文生图 config = GenerationConfig( prompt="A serene Japanese garden with cherry blossoms, ""koi pond, stone bridge, golden hour, cinematic, 8k", negative_prompt="lowres, blurry, text, watermark", width=768, height=512, steps=30, cfg_scale=7.5,) paths = sd.txt2img(config)print(f"生成完成: {paths}")# 图生图if paths: new_paths = sd.img2img( init_image_path=paths[0], prompt="same scene but in autumn, orange and red leaves, snow", denoising_strength=0.6,)print(f"图生图完成: {new_paths}")# 超分辨率if paths: upscaled = sd.upscale(paths[0], scale=2)print(f"超分辨率完成: {upscaled}")# 后处理 pp = ImagePostProcessor()if paths: enhanced = pp.enhance(paths[0]) watermarked = pp.add_watermark(enhanced, text="AI Art")print(f"后处理完成: {watermarked}")# ===== 方式二：云端 API =====# cloud = StabilityAIClient("sk-xxx")# cloud.generate("A futuristic cityscape at sunset", "16:9", "cinematic")if __name__ =="__main__": main()

8. 总结

本文覆盖了 Stable Diffusion 图像生成的完整链路：

  1. 本地部署 SD WebUI 并开启 API 模式
  2. 封装 Python 客户端 支持文生图、图生图、局部重绘、超分辨率
  3. 云端 API 作为无 GPU 环境的替代方案
  4. Prompt 工程 模板化的提示词编写技巧
  5. 后处理管道 增强色彩、添加水印、拼图网格
生成速度参考：RTX 4090 生成 512x512 约 3-5 秒，512x768 约 5-8 秒。云端 API 约 10-20 秒。

Read more

faster-whisper极速安装指南：3分钟搞定AI语音转文字

还在为语音转文字的慢速度而烦恼吗？faster-whisper来拯救你！这款基于OpenAI Whisper模型的优化版本，通过CTranslate2推理引擎实现了4倍速的语音识别，同时保持相同的准确率。无论你是开发者还是技术爱好者，这篇指南将带你轻松上手这个强大的AI语音识别工具。 【免费下载链接】faster-whisper 项目地址: https://gitcode.com/gh_mirrors/fas/faster-whisper 🚀 一分钟快速安装 安装faster-whisper就像呼吸一样简单！只需要一个命令： pip install faster-whisper 是的，就这么简单！Python包管理器会自动处理所有依赖关系，让你在几秒钟内就能开始使用这个强大的语音转文字工具。 VAD语音活动检测模块 - 智能过滤静音片段 🛠️ 硬件环境准备 基础要求 * Python 3.8或更高版本 * 支持CUDA的NVIDIA GPU（推荐）或普通CPU GPU用户专属配置 如果你拥有NVIDIA显卡，为了获得最佳性能，需要安装以下组件：
2026-01-14 学习记录--LLM-申请Hugging Face 访问令牌（以Meta-Llama-3.1-8B-Instruct为例）

2026-01-14 学习记录--LLM-申请Hugging Face 访问令牌（以Meta-Llama-3.1-8B-Instruct为例）

LLM-申请 Hugging Face 访问令牌（以Meta-Llama-3.1-8B-Instruct为例） 一、请求访问Llama模型 ⭐️ 随便进入想要访问的Llama模型，这里展示的是Meta-Llama-3-8B-Instruct。 1、 点击链接，申请访问Llama模型 2、 填写相关申请信息，注意如下：👇🏻(1)、国家最好选「美国」，然后填 「美国的大学」；(2)、操作这一步时，节点需要是对应国家的节点（若是美国，那么节点也要是美国）。 3、 提交成功后，就可开始申请Llama模型的Hugging Face 访问令牌啦~ 二、申请Llama模型的Hugging Face 访问令牌（以Meta-Llama-3.1-8B-Instruct为例）⭐️ 1、判断是否需要申请 访问Meta-Llama-3.1-8B-Instruct模型在Hugging Face上的官方仓库。 假若你看见“You need to agree to share your

Ollama 底层的 llama.cpp 和 GGUF

GGUF = 大模型权重的「通用压缩格式」（类似视频的 MP4，适配所有播放器） llama.cpp = 跑 GGUF 格式模型的「轻量级推理引擎」（类似视频播放器，能在低配电脑上流畅播 MP4） 两者配合：GGUF 让模型体积变小、适配性强，llama.cpp 让模型能在 CPU / 低配 GPU 上快速跑 这也是 Ollama 能做到 “一键本地运行” 的底层原因 GGUF 详解：大模型的 “通用压缩包” 核心定义 GGUF（Generic GGML Format）是 GGML 格式的升级版，是专门为大模型权重设计的二进制存储格式 核心目标是「通用、高效、压缩」 GGML 是什么？
开源浪潮下的中国力量：文心一言大模型本地部署与应用全攻略

开源浪潮下的中国力量：文心一言大模型本地部署与应用全攻略

文章目录 * 一、前言 * 1.1 模型开源意义与背景 * 1.2 文心一言大模型简介 * 1.3 测评目标与思路 * 二、文心一言大模型 * 2.1 文心一言开源概况 * 2.2 文心一言大模型技术综述 * 三、文心一言大模型深度解析 * 3.1 开源策略与生态影响 * 3.1.1 开源时间与版本介绍 * 3.2 模型特性与优势 * 四、部署实战：从 GitCode下载ERNIE-4.5-0.3B 模型到本地可交互服务 * 4.1 环境准备与部署方式 * 4.2 下载与安装步骤 * 4.3 调用示例与接口说明 * 编写部署测试脚本 * 五、