Ask for an outcome, or name the model.
Outcome routes pick the model for you and re-map as models improve — that is the reliable default. Aliases pin a specific model by its canonical name, and most keep multi-provider failover. Preview any row to see the exact URL that invokes it.
Model aliases
Canonical model names, not provider endpoints. Most aliases resolve across several providers for the same model, so pinning one keeps its redundancy.
Google DeepMind
Google DeepMind
Google DeepMind
Google DeepMind
Black Forest Labs
Black Forest Labs
Black Forest Labs
Black Forest Labs
Black Forest Labs
Black Forest Labs
Black Forest Labs
Black Forest Labs
Black Forest Labs
Recraft
Ideogram
Alibaba Qwen
ByteDance Seed
Z.ai
xAI
RunDiffusion
Vyro
VyroAn alias we do not recognise is never an error — the request falls back to quality/fast and says so in an IGPT-Warning header. 4 older spellings — gemini-3-pro, nano-banana-2, nano-banana-2-lite, gemini-2.5-flash — still resolve so existing URLs keep their cache, but they are not listed here.
Reference images: 14 of these 22 aliases accept an image_url — gemini-3-pro-image, gemini-3.1-flash-image, gemini-3.1-flash-lite-image, gemini-2.5-flash-image, flux-2-dev, flux-2-pro, flux-2-dev-turbo, flux-2-klein-4b, flux-2-klein-4b-distilled, flux-2-klein-9b, flux-1-schnell, seedream-4.5, glm-image, juggernaut-flux-pro. On the rest it is dropped and the prompt alone is generated.
Routes
Three outcomes at three levels. Omit route and you get quality/fast.
How a route gets its model
We score models against the thing the route promises — prompt adherence, legible type, photographic realism — and at the level you asked for. The winner serves the route until it is beaten.
Adding a model is not your migration
New models appear behind existing outcome routes. Nothing in your codebase changes, no URL breaks, and already-cached images keep serving exactly as generated.
Reference images and presets
Supported models accept an image_url for grounding and style. Save a prompt and parameter set as a preset to keep generations predictable across a project.