GPT Image 2 Multi-Model Image Generation

Problem

AI agents often lack a stable image-generation entry point. API keys may come from environment variables or host configuration, while endpoints, models, sizes, and response formats vary. Hard-coding one provider makes the workflow brittle across gateways, compatible APIs, or local models. This skill turns image generation into a reusable agent capability with a consistent discovery and execution path.

How It Works

• API discovery: reads environment variables first, then host files such as config.yaml or settings.json to find an image endpoint.
• Model selection: prefers gpt-image-2, then dall-e-3, dalle-3, stable-diffusion-xl, or any available model, with endpoint validation.
• Structured prompts: assembles Subject / Action / Context / Composition / Lighting / Style to reduce unstable outputs from vague prompts.
• Generation and saving: calls /v1/images/generations, handles b64_json or url responses, and writes images to a local directory.
• Recovery and memory: maps errors such as 401, 403, 429, and timeout, retries with exponential backoff, and updates session notes.

Boundaries

It works best when an OpenAI-compatible image API is already configured and the agent needs to pick models automatically. It will still fail if no usable API exists or the endpoint does not support image generation. Keep model size limits in mind: gpt-image-2 supports 1024x1024 and 2048x1024, while dall-e-3 commonly supports 1024x1024 and 1792x1024.