Image generation¶

Text → image, tri-parity across Azure / GCP / AWS. Same discipline as every earlier recipe: verified model IDs, one dispatch function, deploy composes E1.

The pipeline¶

     ┌────────────┐    ┌──────────────┐    ┌─────────────┐
     │  prompt    │──► │  text→image  │──► │ image bytes │
     │  (text)    │    │  (per cloud) │    │  (PNG/JPEG) │
     └────────────┘    └──────────────┘    └─────────────┘

No LLM in the middle. The point here is the model catalog + request shape per cloud.

Per-cloud services¶

Cloud	Service	Default model	Package
Azure	Azure OpenAI Images	`gpt-image-1`	`openai` (same SDK as chat)
GCP	Vertex AI Imagen	`imagen-4.0-generate-001`	`google-genai` (same SDK as chat)
AWS	Bedrock — Amazon Nova Canvas	`amazon.nova-canvas-v1:0`	`boto3` `bedrock-runtime`

The GCP and Azure paths reuse the exact SDK you set up in Call an LLM — only the method changes (images.generate on Azure, models.generate_images on Vertex). AWS is a different Bedrock model id but the same invoke_model shape as embeddings.

Docs verified 2026-08-08¶

Azure OpenAI image generation — learn.microsoft.com/…/openai/how-to/dall-e
Vertex Imagen — ai.google.dev/gemini-api/docs/imagen
Bedrock Nova Canvas overview — docs.aws.amazon.com/nova/…/image-generation
Bedrock Nova Canvas request/response — docs.aws.amazon.com/nova/…/image-gen-req-resp-structure

Model catalog snapshot¶

Provider	Model IDs (verified)	Notes
Azure OpenAI	`gpt-image-2` (GA), `gpt-image-1.5`, `gpt-image-1`, `gpt-image-1-mini`	DALL-E-3 retired 2026-03-04. Sizes: 1024×1024, 1024×1536, 1536×1024 (GPT-Image-1 family); GPT-Image-2 accepts arbitrary sizes with 16-px multiple + long edge ≤ 3840 px.
Vertex Imagen	`imagen-4.0-generate-001`, `imagen-4.0-fast-generate-001`, `imagen-4.0-ultra-generate-001`, `imagen-3.0-generate-002`	Aspect ratios: `1:1`, `3:4`, `4:3`, `9:16`, `16:9`. Imagen shuts down 2026-08-17 on the Gemini Developer API — Nano Banana 2 (`gemini-3.1-flash-image`) is the current migration target; Nano Banana Pro (`gemini-3-pro-image`) is the studio-quality tier. Original Nano Banana (`gemini-2.5-flash-image`) remains GA.
Bedrock	`amazon.nova-canvas-v1:0`	Nova Canvas covers T2I + editing (inpainting/outpainting/background removal/virtual try-on/color-guided/image conditioning/image variation). Max 2048×2048 (or 2816×1536) for generation.

Pages¶

Generate — one call per cloud, verified request shapes, returns image bytes.
Deploy — Terraform composing E1 baseline + per-cloud image-service resources (IAM for Bedrock Nova Canvas; no per-cloud API enable needed for Azure OpenAI / Vertex Imagen since they run on the same accounts as chat).

Validate-only¶

Same discipline as every prior batch — no live image generation in CI. The "verify" block on the Generate page has the local one-liner.