Image Generation Capability Integration

← Developer Docs · 简体中文

Is Code Required?

Always required. The image generation API has no industry-standard format. Each vendor uses different proprietary API formats, so regardless of whether the platform is OpenAI-compatible for chat, you need to write a Service class.

Prerequisite: Register Platform Name Constant

Add a new platform name constant in AdiConstant.ModelPlatform:

File: adi-common/src/main/java/com/moyz/adi/common/cosntant/AdiConstant.java

public static class ModelPlatform {
    // ... existing constants ...
    public static final String NEWAI = "newai";
}

Step 1: Service Development

Inheritance: CommonModelService → AbstractImageModelService → Your Service

The base class AbstractImageModelService implements the complete image generation flow (calling the model, handling URL/Base64 responses, saving files). Subclasses only need to provide an ImageModel instance and handle errors.

Required abstract methods:

Method	Return Type	Description
isEnabled()	boolean	Check if the service is available
buildImageModel(User, Draw)	ImageModel	Build the image generation model
parseError(Object)	LLMException	Parse API errors

Complexity Notes

The actual development complexity depends on how different the target platform's API is:

• Platforms already supported by langchain4j (e.g. OpenAI): Use OpenAiImageModel directly — very simple (see OpenAiImageService; it dispatches t2i vs i2i: without reference images it uses the built-in model, with them the custom OpenAiImageEditModel calling /images/edits)
• Platforms requiring custom adaptation: You need to write a custom ImageModel implementation to interface with the platform API, then use it in the Service (see DashScopeImageService; t2i and i2i share one endpoint, reference images are relayed as base64 data URIs via AdiImageGenerationModel)

Step 2: Register in AiModelInitializer

File: adi-common/src/main/java/com/moyz/adi/common/service/AiModelInitializer.java

Add in the initImageModelServiceList() method:

private synchronized void initImageModelServiceList(Map<String, ModelPlatform> nameToPlatform) {
    // ... existing platforms ...

    // Add new
    initImageModelService(AdiConstant.ModelPlatform.NEWAI,
        model -> new NewAiImageService(model, nameToPlatform.get(AdiConstant.ModelPlatform.NEWAI)));
}

Step 3: Frontend Adaptation (User Web)

File: langchain4j-aideepin-web/src/views/draw/

The frontend uses a single unified form component GenerateImageForm.vue; form fields (size/count/quality/seed/reference images) are all driven by the model's properties, so new platforms usually need no new frontend component.

Frontend Component Structure

src/views/draw/components/
  ├── GenerateImageForm.vue      — Unified generation form (shared by all platforms; fields rendered per model properties)
  ├── RefImageUpload.vue         — Reference-image upload for i2i (shown for models whose input_types include image)
  ├── ParamsPanel.vue            — Parameter section collapse/expand (floating arrow)
  ├── CommonDraws.vue            — Common draw operations
  ├── Header.vue                 — Page header (with model selector)
  └── SearchInput.vue            — Prompt input

What to Develop

Usually no frontend work: all platforms share the unified GenerateImageForm.vue, whose fields (size/count/quality/seed/reference images) render from the model's adi_ai_model.properties declarations. Integrating a new platform only requires configuring the model in the admin console with (all five keys are declared uniformly; empty/false/1 hides the corresponding form field):

properties key	Purpose	When empty
sizes	Size options ([{value,label}])	Empty array falls back to defaults
max_images	Per-run image count cap (slider shown when > 1)	1 (no slider)
qualities	Quality options (e.g. ["auto","low","medium","high"])	No quality field
is_support_seed	Whether the seed field is shown	false hides the seed field
max_ref_images	Reference-image limit (paired with image in input_types)	1

The reference-image upload appears automatically for models whose input_types include image — no frontend change needed.

---

Previous: Chat / Vision Integration · Next: TTS Integration