AWS Bedrock Image Generation Setup¶

Overview¶

Replace non-AWS image generators (Midjourney, FalAI, DALL-E direct) with AWS Bedrock Stable Diffusion SDXL for 100% AWS compliance.

Current Status¶

✅ Already Integrated¶

AWS Bedrock SDXL driver exists: app/Domains/Entity/Drivers/StableDiffusion/AwsBedrockDriver.php
BedrockEngine enum configured: app/Enums/BedrockEngine.php
Model available: STABLE_DIFFUSION_1 = stability.stable-diffusion-xl-v1

❌ Removed (Non-Compliant)¶

Midjourney (via PiAPI)
FalAI (Kling, Veo, Flux Pro, Ideogram, etc.)
Creatify (ad marketing videos)
Standalone image generators

⚠️ Still Available (Other AWS Services)¶

OpenAI DALL-E - If using via AWS integration
Together Flux - If hosted on AWS infrastructure

AWS Bedrock Image Generation¶

Supported Model¶

Stable Diffusion XL (SDXL) - Model ID: stability.stable-diffusion-xl-v1 - Resolution: 1024x1024, 512x512, or custom - Speed: ~20-30 seconds per image - Cost: ~$0.04-0.05 per image

Features¶

✅ Text-to-Image - Generate images from text prompts - Support for style presets - Negative prompts for quality control

✅ Image-to-Image - Modify existing images - Style transfer - Upscaling

✅ Inpainting - Edit specific parts of images - Remove/replace objects

Implementation¶

1. Enable AWS Bedrock SDXL¶

The driver is already implemented. You need to:

Enable the model in AWS Console:

AWS Console → Bedrock → Model access → Request access
Select: Stable Diffusion XL 1.0
Wait for approval (~5-10 minutes)

Configure in your app: Already configured in app/Enums/BedrockEngine.php:
```
case STABLE_DIFFUSION_1 = 'stability.stable-diffusion-xl-v1';
```

2. Usage Example¶

Generate Image via Bedrock:

use App\Domains\Entity\Enums\EntityEnum;
use App\Domains\Entity\Facades\Entity;

// Generate image
$result = Entity::driver(EntityEnum::STABLE_DIFFUSION_XL_1024_V_1_0)
    ->input([
        'prompt' => 'A beautiful sunset over mountains, professional photography',
        'negative_prompt' => 'blurry, low quality, watermark',
        'width' => 1024,
        'height' => 1024,
        'steps' => 30,
        'cfg_scale' => 7.0,
        'samples' => 1,
    ])
    ->generate();

// Save to S3
if ($result->success) {
    $imageData = base64_decode($result->data['artifacts'][0]['base64']);
    $filename = 'generated_' . time() . '.png';
    Storage::disk('s3')->put("images/{$filename}", $imageData, 'public');
    $url = Storage::disk('s3')->url("images/{$filename}");
}

3. Available Entities¶

Check EntityEnum.php for Stable Diffusion variants: - STABLE_DIFFUSION_XL_1024_V_1_0 - Main SDXL model - STABLE_DIFFUSION_512_V_2_1 - Faster, smaller images - IMAGE_TO_VIDEO - Convert images to video - Other Stable Diffusion variants

Migrating Existing Images from R2 to S3¶

Migration Script¶

A migration is included: database/migrations/2025_11_21_000002_migrate_r2_images_to_s3.php

Run migration:

# First, ensure R2 credentials are still in .env (temporarily)
R2_ACCESS_KEY_ID=...
R2_SECRET_ACCESS_KEY=...
R2_BUCKET=...
R2_ENDPOINT=...

# Run migration
php artisan migrate

# After migration completes, remove R2 credentials from .env

What it does: 1. Scans database for R2 image URLs 2. Downloads images from R2 3. Uploads to AWS S3 4. Updates database URLs to point to S3 5. Keeps original R2 files as backup

Tables checked: - user_openai (output, image columns) - user_openai_chat_messages (output) - ai_presentations (pdf_url, pptx_url)

Image Generator Replacement Options¶

Option 1: Use AWS Bedrock SDXL (Recommended)¶

Update your image generation UI:

// OLD (Non-compliant)
$engines = [
    'dalle2' => 'DALL-E 2',
    'dalle3' => 'DALL-E 3',
    'midjourney' => 'Midjourney',
    'stable-diffusion' => 'Stable Diffusion',
];

// NEW (AWS-compliant)
$engines = [
    'aws-bedrock-sdxl' => 'AWS Bedrock SDXL',
    'stable-diffusion' => 'Stable Diffusion (Legacy)',
];

Controller update:

// In AIController or similar
public function generateImage(Request $request)
{
    // Use AWS Bedrock SDXL
    $entity = EntityEnum::STABLE_DIFFUSION_XL_1024_V_1_0;

    $result = Entity::driver($entity)
        ->input([
            'prompt' => $request->prompt,
            'negative_prompt' => $request->negative_prompt ?? 'low quality, blurry',
            'width' => $request->width ?? 1024,
            'height' => $request->height ?? 1024,
            'steps' => 30,
            'cfg_scale' => 7.0,
        ])
        ->generate();

    // Handle result...
}

Option 2: Keep OpenAI DALL-E (If AWS-Hosted)¶

If you're using OpenAI via an AWS-hosted integration, you can keep it: - Make sure OpenAI API calls go through AWS infrastructure - Document the AWS hosting for compliance

Option 3: Self-Host Open Source Models on AWS¶

For maximum control: - Deploy Stable Diffusion on AWS SageMaker - Use AWS EC2 with GPU instances - More expensive but fully controlled

Cost Comparison¶

Service	Cost per Image	Compliance
Midjourney	$0.03-0.10	❌ Non-AWS
FalAI (Flux, etc.)	$0.04-0.08	❌ Non-AWS
AWS Bedrock SDXL	$0.04-0.05	✅ AWS-compliant
Self-hosted on SageMaker	$0.10-0.20	✅ AWS-compliant

Verdict: AWS Bedrock SDXL has competitive pricing and is fully compliant.

Style Presets (SDXL)¶

AWS Bedrock SDXL supports these style presets:

photographic - Photorealistic images
digital-art - Digital artwork style
cinematic - Movie/film style
anime - Anime/manga style
line-art - Line drawing style
3d-model - 3D rendered style
analog-film - Vintage film look
fantasy-art - Fantasy illustration
isometric - Isometric 3D view
low-poly - Low polygon 3D art
modeling-compound - Clay/sculpture style
neon-punk - Cyberpunk neon style
origami - Paper folding style
pixel-art - Retro pixel art

Usage:

$result = Entity::driver($entity)
    ->input([
        'prompt' => 'A futuristic city',
        'style_preset' => 'neon-punk',
        // ... other params
    ])
    ->generate();

Quality Tips¶

Better Prompts¶

Good:

"Professional portrait photo of a business executive,
studio lighting, sharp focus, 4k quality, photorealistic"

Bad:

"person in suit"

Negative Prompts¶

Always include to improve quality:

"blurry, low quality, distorted, watermark, text,
cropped, out of frame, bad anatomy"

Recommended Settings¶

For best quality: - Steps: 30-50 - CFG Scale: 7-10 - Resolution: 1024x1024

For speed: - Steps: 20-30 - CFG Scale: 7 - Resolution: 512x512

Testing¶

Test Image Generation¶

php artisan tinker

use App\Domains\Entity\Enums\EntityEnum;
use App\Domains\Entity\Facades\Entity;

$result = Entity::driver(EntityEnum::STABLE_DIFFUSION_XL_1024_V_1_0)
    ->input([
        'prompt' => 'A majestic lion in the savanna, golden hour lighting, professional wildlife photography',
        'negative_prompt' => 'blurry, low quality',
        'width' => 1024,
        'height' => 1024,
    ])
    ->generate();

dump($result);

Verify S3 Storage¶

aws s3 ls s3://your-bucket/images/ --recursive

Troubleshooting¶

Error: "Model not enabled"¶

Solution: Request access to SDXL in AWS Bedrock console

Error: "Insufficient permissions"¶

Solution: Add Bedrock permissions to IAM role:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:InvokeModel",
        "bedrock:InvokeModelWithResponseStream"
      ],
      "Resource": "arn:aws:bedrock:*::foundation-model/stability.stable-diffusion-xl-v1"
    }
  ]
}

Images are blurry¶

Solution: Increase steps and resolution:

'steps' => 50,
'width' => 1024,
'height' => 1024,
'cfg_scale' => 8.0,

Migration Checklist¶

Support¶

AWS Bedrock Docs: https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-diffusion.html Stable Diffusion Guide: https://platform.stability.ai/docs/api-reference

Issues: https://github.com/vell-io/vell-main/issues

Last Updated: 2025-11-21 Compliance Status: ✅ AWS-Compliant