Stable Diffusion与PHP结合：打造强大的本地部署系统

引言

Stable Diffusion作为当前最热门的AI图像生成模型之一，让很多开发者都想将其集成到自己的应用中。本文将详细介绍如何将Stable Diffusion与PHP结合，构建一个完整的本地部署系统。通过这种方式，你可以创建自己的AI图像生成网站或API服务。

准备工作

在开始之前，请确保你的系统满足以下要求：

硬件要求：
- 至少8GB显存的NVIDIA显卡（推荐RTX 3060及以上）
- 16GB以上内存
- 50GB以上可用磁盘空间
软件要求：
- Python 3.10+
- PHP 7.4+
- Composer (PHP依赖管理工具)
- Git

第一部分：安装Stable Diffusion

1.1 克隆Stable Diffusion WebUI仓库

代码片段

git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
cd stable-diffusion-webui

1.2 安装依赖项

代码片段

python -m venv venv
source venv/bin/activate  # Linux/MacOS
# Windows使用: venv\Scripts\activate

pip install torch torchvision --extra-index-url https://download.pytorch.org/whl/cu117
pip install -r requirements.txt

1.3 下载模型文件

将Stable Diffusion模型文件(.ckpt或.safetensors)放入models/Stable-diffusion目录。

1.4 启动WebUI测试

代码片段

python launch.py --listen --api

参数说明：
– --listen: 允许网络访问
– --api: 启用API接口

访问http://localhost:7860确认安装成功。

第二部分：PHP与Stable Diffusion API集成

2.1 PHP环境准备

创建一个新项目目录并初始化Composer：

代码片段

mkdir sd-php-integration
cd sd-php-integration
composer init --no-interaction
composer require guzzlehttp/guzzle

2.2 PHP调用SD API的示例代码

创建generate.php文件：

代码片段

<?php
require 'vendor/autoload.php';

use GuzzleHttp\Client;
use GuzzleHttp\Exception\RequestException;

class StableDiffusionAPI {
    private $client;
    private $apiUrl = 'http://localhost:7860/sdapi/v1/txt2img';

    public function __construct() {
        $this->client = new Client([
            'timeout' => 60.0, // AI生成可能需要较长时间
        ]);
    }

    /**
     * Generate image from text prompt
     * 
     * @param string $prompt The text prompt for image generation
     * @param array $options Additional generation options
     * @return array Contains 'image' (base64) and 'info' (generation parameters)
     */
    public function generateImage($prompt, $options = []) {
        $defaultOptions = [
            'prompt' => $prompt,
            'negative_prompt' => '', // Negative prompts to avoid certain elements
            'steps' => 20,          // Number of diffusion steps (20-50 is typical)
            'width' => 512,         // Image width (must be multiple of 64)
            'height' => 512,        // Image height (must be multiple of 64)
            'cfg_scale' => 7,       // Classifier-free guidance scale (7 is a good balance)
            'sampler_name' => 'Euler a', // Sampling method ('Euler a', 'DPM++ SDE Karras', etc.)
            'seed' => -1,           // Random seed (-1 for random)
        ];

        $payload = array_merge($defaultOptions, $options);

        try {
            $response = $this->client->post($this->apiUrl, [
                'json' => $payload,
                'headers' => [
                    'Content-Type' => 'application/json',
                ]
            ]);

            return json_decode($response->getBody(), true);

        } catch (RequestException $e) {
            throw new Exception("API request failed: " . $e->getMessage());
        }
    }

    /**
     * Save base64 image to file
     */
    public function saveImage($base64Data, $filename) {
        if (preg_match('/^data:image\/(\w+);base64,/', $base64Data)) {
            $base64Data = substr($base64Data, strpos($base64Data, ',') + 1);
        }

        file_put_contents($filename, base64_decode($base64Data));
    }
}

// Usage example:
try {
    $sd = new StableDiffusionAPI();

    // Generate an image with custom parameters
    $result = $sd->generateImage(
        "a beautiful sunset over mountains, digital art", 
        [
            'steps' => 25,
            'width' => 768,
            'height' => 512,
            'sampler_name' => 'DPM++ SDE Karras'
        ]
    );

    // Save the generated image
    if (!empty($result['images'][0])) {
        $sd->saveImage($result['images'][0], 'generated_image.png');
        echo "Image generated and saved successfully!";

        // Output generation info for debugging
        echo "<pre>Generation info:\n" . print_r(json_decode($result['info'], true), true) . "</pre>";
    } else {
        echo "No image was generated.";
    }

} catch (Exception $e) {
    echo "Error: " . $e->getMessage();
}
?>

API参数详解

在调用API时，有几个关键参数需要注意：

Prompt工程：
代码片段
```
"a beautiful sunset over mountains, digital art"
```
- Prompt质量直接影响生成结果质量。建议：
  - 具体描述：不要只说”一只猫”，而是说”一只橘色的短毛猫坐在窗台上，阳光照射在毛发上”
  - 风格修饰：添加如”digital art”, “photorealistic”, “oil painting”等风格词
负面Prompt：
代码片段
```
"negative_prompt": "blurry, low quality, distorted"
```
- AI不知道什么是”不好”，需要明确告诉它避免什么
采样方法：
代码片段
```
"sampler_name": "DPM++ SDE Karras"
```
- Euler a：快速但可能缺乏细节（适合草图）
- DPM++ SDE Karras：高质量但较慢（适合最终作品）
CFG Scale：
代码片段
```
"cfg_scale":7 
```
- 3-5：创意性强但可能偏离Prompt描述较多
- 7-10：平衡创意和Prompt遵循度
- >15：严格遵循Prompt但可能失去自然感
Steps步数：
代码片段
```
"steps":25 
```
- 20-30步通常足够
- >50步通常不会带来明显质量提升
分辨率选择：
代码片段
```
"width":768,"height":512 
```
- SD v1.x系列模型建议不超过768×768
- SDXL模型可支持1024×1024

PHP前端交互示例（可选）

如果你想创建一个简单的Web界面来测试你的API，可以添加以下HTML表单：

创建index.html：

代码片段

<!DOCTYPE html>
<html>
<head>
    <title>Stable Diffusion PHP Demo</title>
</head>
<body>
    <h1>AI Image Generator</h1>

    <form action="generate.php" method="post">
        <div>
            <label for="prompt">Prompt:</label><br>
            <textarea name="prompt" rows="4" cols="50" required></textarea>
        </div>

        <div>
            <label for="negative_prompt">Negative Prompt:</label><br>
            <textarea name="negative_prompt" rows="2" cols="50"></textarea>
        </div>

        <div>
            <label for="steps">Steps:</label>
            <input type="number" name="steps" value="20" min="10" max="150">

            <label for="width">Width:</label>
            <input type="number" name="width" value="512" min="256" max="1024">

            <label for="height">Height:</label>  
            <input type="number" name="height" value="512" min="256" max="1024">

            <label for="cfg_scale">CFG Scale:</label>  
            <input type="number" step=0.5 name=cfg_scale value=7 min=3 max=20>  

             <!-- Sampler selection -->
             <select name=sampler_name required>   
                 <?php foreach(['Euler a','DPM++ SDE Karras','LMS','Heun'] as$sampler): ?>
                     <option value="<?=$sampler?>"><?=$sampler?></option>   
                 <?php endforeach; ?>   
             </select>   
         </div>   

         <!-- Seed control -->   
         <div style=margin-top:10px>   
             Seed (-1 for random):   
             <input type=number name=seed value=-1 min=-1 max=9999999999 step=1>   
         </div>   

         <!-- Submit button -->   
         <button type=submit style=margin-top:15px;padding:8px16px;background:#007bff;color:white;border:none;border-radius:4px;cursor:pointer>Generate Image</button>     
     </form>     

     <?php if(isset($_GET['image'])): ?>     
         <!-- Display generated image if available -->     
         <?php$imageFile='generated/'.basename($_GET['image']);if(file_exists($imageFile)):?>     
             <?phplist($width,$height)=getimagesize($imageFile);?>     
             <?phpecho "<p><strong>{$width}x{$height}</strong></p>"?>     
             <?phpecho "<img src='{$imageFile}'style='max-width:100%'>"?>     
         <?phpendif?>     
     <?phpendif?>     
 </body></html>

修改后的完整PHP处理脚本:

代码片段


<?php 

require_once __DIR__.'/vendor/autoload.php'; 

use GuzzleHttp\Client; 

// Ensure output directory exists  
if(!is_dir(__DIR__.'/generated')){mkdir(__DIR__.'/generated');}  

// Handle form submission  
if($_SERVER['REQUEST_METHOD']==='POST'){  

$client=new Client([/*...*/]);  

$payload=[/*...*/];  

try{  

$response=$client->post('http://localhost:7860/sdapi/v1/txt2img',[/*...*/]);  

$data=json_decode((string)$response->getBody(),true);  

if(!empty($data['images'][0])){  

// Generate unique filename  
$filename='sd_'.time().'.png';  

// Save the image  
file_put_contents(__DIR__."/generated/{$filename}",base64_decode(preg_replace('#^data.*?base64,#','',$data['images'][0])));  

// Redirect to display the result  
header("Location:/?image={$filename}");exit();}else{thrownewException('No images were generated');}}catch(Exception$e){die("Error generating image:" . htmlspecialchars($e->getMessage()));}}?>

PHP高级集成技巧

当你在生产环境中部署时，需要考虑以下几个高级功能:

队列系统

长时间运行的AI任务应该使用队列系统来处理:

代码片段

composer require enqueue/enqueue enqueue/fs # Simple filesystem queue

然后修改生成代码:

代码片段


<?php 

use Enqueue\Fs\FsConnectionFactory;

// In your controller action:
function generateAction(Request$request){

// Push job to queue instead of processing immediately

$connectionFactory=new FsConnectionFactory('file://'.__DIR__.'/queue');

$context=$connectionFactory->createContext();

$queue=$context->createQueue('sd_generation');

$message=$context->createMessage(json_encode([

'prompt'=>$request->get('prompt'),

// Other params...

]));

$context->createProducer()->send($queue,$message);

return new Response('Your image is being generated...');}

Worker脚本 (worker.php) :

代码片段


<?php 

require __DIR__.'/vendor/autoload.php';

use Enqueue\Fs\FsConnectionFactory;

while(true){

try{

$connectionFactory=new FsConnectionFactory('file://'.__DIR__.'/queue');

$context=$connectionFactory->createContext();

$queue=$context->createQueue('sd_generation');

if(null!== ($message=$context->createConsumer ($queue)->receive(5000))){ /* Wait up to5 sec */

echo "\nProcessing job...";

processJob(json_decode ($message -> getBody (),true));

echo "\nDone.";

/* Acknowledge the message */

return;
}}catch(Exception){sleep(5);continue;}}
function processJob(array){
/* Your existing generation logic here */}?>

运行worker进程:

代码片段

nohup php worker.php > worker.log & # Run in background on Linux 
Start-Process php-FilePath worker.php-WindowStyle Hidden # On Windows

性能优化

对于高流量站点，可以实施以下优化策略:

Caching层

缓存常见提示词的结果:

代码片段

composer require symfony/cache

使用示例:

代码片段


<?php 

use Symfony\Component\Cache\Adapter\FilesystemAdapter;

function getCachedGeneration(string){

/* Create cache pool */

static$cache=new FilesystemAdapter('app.cache',3600,'./cache');

/* Generate cache key from prompt + params */

return;}

Batch Processing

当需要生成多张图片时(如电商产品变体)，使用批处理模式更高效:

修改SD API调用为批处理模式:

代码片段


-$payload=[
+$payload=[
+      batch_size=>3,
       prompt=>[
+           prompt=>["a cat","a dog","a bird"],
+           negative_prompt=>["ugly","low quality"]*3,
       ],
       ...
];

然后在PHP中并行处理:

代码片段

composer require spatie/guzzle-rate-limiter-middleware spatie/fork ^3.0 --dev # For parallel processing in dev env only! Not recommended for production!

并行处理示例(仅限开发环境):

(点击展开)

 fork(function()use ($p){ return generateSingleImage ($p); }),)); } /* Helper function */ function generateSingleImage(array){ /* Your existing single-image generation logic */ } ?>

在生产环境中应该使用专业队列系统如RabbitMQ或AWS SQS.

GPU资源管理

当多个用户同时请求时，需要管理GPU资源以避免OOM错误:

实现简单的令牌桶算法来控制并发请求数:

(点击展开)

class GpuResourceManager { private int ; private int ; private float ; public function __construct(int ,int ){ /* Initialize token bucket */ } public function acquireToken():bool { /* Implement token bucket algorithm */ } }

/* Usage example : */

function generateWithResourceManagement(){ global ; if (! -> acquireToken()){ throw new Exception ('Server busy , please try again later'); } try { return generateImage(); } finally { -> releaseToken(); } }

Monitoring & Logging

添加监控以跟踪GPU使用情况 :

(点击展开)

Install Prometheus client library composer require prometheus/client_guzzle ^6 . x-dev composer require prometheus/prometheus ^2 . x-dev # Configure metrics endpoint route (' / metrics ',function (){ header (' Content-Type : text / plain ; version =0 .0 .4'); echo Registry :: getDefaultRegistry () -> export (); });

然后在Nginx配置中添加 :

location ~ ^ / metrics { allow YOURMONITORINGSERVER_IP ; deny all ; }

最后设置Grafana仪表板来可视化 :

• GPU显存使用率 • API响应时间 • Queue积压情况 • Error rates等关键指标.

Security Considerations安全考虑 ### Authentication & Rate Limiting Always protect your API endpoints : composer require symfony / rate-limiter Then implement rate limiting middleware : class RateLimiterMiddleware { public function handle ($ request , Closure ){ if (! app () -> make (\ Illuminate \ Contracts \ Cache :: class ) -> tooManyAttempts (” ip :”. request () ip (), perMinute :)) abort (, message : ); return next (); }} Route middleware ([ ])-> group (function (){});

For user authentication , consider Laravel Sanctum or JWT .

Input Sanitization Always sanitize user inputs before sending to SD API : use voku / anti-xss ; function sanitizePrompt(string ):string { static xssCleaner =new AntiXSS(); return xssCleaner -> xssclean( striptags());}

This prevents prompt injection attacks where malicious users might try to inject HTML / JavaScript into your system through cleverly crafted prompts .

Troubleshooting常见问题解决 ### CUDA Out Of Memory Errors If you encounter CUDA OOM errors , try these solutions : • Reduce batch_size in API calls • Lower resolution (e.g., from512x512to384x384) • Enable