AI Image Generator Bundle

This is bundle utalizes common APIs for generative image AIs to generate images in the Pimcore Backend.

Installation

composer update basilicom/ai-image-generator-bundle

Make sure to also install the bundle via BundleSetupSubscriber or console.

Support

Configuration

ai_image_generator:
   brand:
    colors:
      - "#0062FF"
      - "#B34197"
      - "#FF444A"

  prompt_enhancement:
    service:        ~|ollama|basilicom|open_ai

    services:
      ollama:
        baseUrl:    "http://localhost:11434/"
        model:      "llama2"

      basilicom:
        baseUrl:    "http://localhost:8080/"

      open_ai:
        baseUrl:    "https://api.openai.com/v1"
        apiKey:     "%env(OPEN_AI_API_KEY)%"

  feature_services:
    txt2img:            open_ai | stable_diffusion_api | dream_studio | clip_drop
    image_variations:   open_ai | stable_diffusion_api | dream_studio | clip_drop
    upscale:            -       | stable_diffusion_api | dream_studio | clip_drop
    inpaint:            open_ai | stable_diffusion_api | dream_studio | -
    inpaint_background: open_ai | stable_diffusion_api | -            | clip_drop

  services:
    stable_diffusion_api:
      baseUrl:        "http://host.docker.internal:7860"
      model:          "JuggernautXL"
      inpaint_model:  "JuggernautXL"
      steps:          30
      upscaler:       "ESRGAN_4x"

    dream_studio:
      baseUrl:        "https://api.stability.ai"
      model:          "stable-diffusion-xl-beta-v2-2-2"
      inpaint_model:  "stable-diffusion-xl-1024-v1-0"
      steps:          10 
      apiKey:         "%env(DREAM_STUDIO_API_KEY)%"
      upscaler:       "esrgan-v1-x2plus"

    open_ai:
      baseUrl:        "https://api.openai.com/v1"
      apiKey:         "%env(OPEN_AI_API_KEY)%"

    clip_drop:
      baseUrl:        "https://clipdrop-api.co"
      apiKey:         "%env(CLIP_DROP_API_KEY)%"

Usage

Generating images in documents

If no prompt is given, the prompt will be generated (and not translated!) from

• document SEO title
• document SEO description
• h1-Elements
• h2-Elements
• h3- and h4-elements if the previous mentioned sources are empty

Generating images in DataObjects

If no prompt is given, the prompt will be generated (and not translated!) by trying to access the following properties:

• key
• title
• name
• productName
• description

API

(POST) /admin/ai-images/generate/{context}-{id}

Generate an image based on a document or object context. If the prompt is empty, the budle-logic for prompting will take effect.

(POST) /admin/ai-images/upscale/{id}

Upscale image, while the target upscaling size is AI-Service specific

(POST) /admin/ai-images/vary/{id}

Inpaint backgrounds where the background logic differ for provided AI-Services.

Responses

Based on the Accept-header, you can say if you want to have a JSON-response or the image itself.

Accept: application/json

{
  success: true,
  id: Pimcore-Asset-ID,
  image: "base64-decoded Image",   
}

{
  success: false,
  message: "..."
}

Accept: image/jpeg

// the base64 decoded image

Using Stable Diffusion API

When running Automatic1111 locally, you can define http://host.docker.internal:7860 as your local API-url.

Additionally, make sure you started Automatic1111 with --api:

  ./webui.sh --api # windows
  ./webui.bat --api # linux/mac

If you want to know which models you have, call the Models-Endpoint and copy the name of a model of your choice.

Plugins used

• ControlNet with canny and ip2p
• SD Upscaler Post Processor Script

Using LLM-driven prompt enhancing

In order to enhance prompts, we use local images of LLMs. There are three supported prompt enhancement services:

• open_ai (ChatGPT)
• basilicom (a simple LLM implementation, see Docker Hub)
• ollama (see Github)

Limitations

• the DreamStudio REST API does currently not support variations, I'll look forward to use the gRPC API
• ClipDrops text-to-image API can only create 1:1 images

Additional ideas

• Prompting
  • enhance prompts, especially for background inpainting, like

    background = "a creepy forest at night"
    image_type = "a haunted castle background"
    characters = "medieval warriors"
    action = "fighting for the honor"
    prompt = f"{image_type} in {background} with {characters} {action}"

• generate prompt in lightbox before sending?
• background-inpainting for other service by using masks
• CLIP interrogate in order to optimize variation prompting
  • allow variants by img2img and CLIP
• run IMG2IMG with low denoise on background-inpainting
• LCM for super fast preview generation => midjourney-like/inpainting-like image selection before upscaling, etc.
• outpainting via Thumbnail
• better error handling (warnings and fallbacks if credits exceeded)
• ComfyUI + Nodes to Python as fixed presets
  • allow docker images with presets
• InvokeAI

Authors

Alexander Heidrich