Image generation

Image models (gpt-image-1, DALL·E 3, FLUX and others with an OpenAI-compatible API) are served at POST /v1/images/generations (a JSON body) and POST /v1/images/edits (a multipart/form-data upload of one or more pictures in image or image[], and an optional mask). Clients use the OpenAI SDKs unchanged:

curl https://<gateway>/v1/images/generations -H "Authorization: Bearer $UME_API_KEY" \
  -H "Content-Type: application/json" -d '{"model":"ume/image","prompt":"En röd stuga vid en sjö","size":"1024x1024"}'
curl https://<gateway>/v1/images/edits -H "Authorization: Bearer $UME_API_KEY" \
  -F model=ume/image -F prompt="Gör himlen blå" -F image=@stuga.png

Setting it up. OpenAI, Azure OpenAI and Foundry providers serve images; an OpenAI-compatible server needs Image generation and editing under Endpoints this server offers. Add the model with kind Image generation (Image; discovery does this for gpt-image*, dall-e*, flux*, imagen* and stable-diffusion* models) and put it behind an alias with the same kind. Fallback, residency, provider allow-lists, rate limits, budgets and routing rules (endpoint == "image_generations" or "image_edits") work as for chat. The dev seed has ume/image on FakeLlm.

Prices. DALL·E-style models are priced per picture (Image (USD per picture) in the price form, perImageUsd); gpt-image-1 per token (the usual input and output prices; the answer's usage is read). The pictures are counted from the answer's data array, so a request for n=4 that returns three is billed for three. A per-picture price is one flat amount for every size and quality, so it is an estimate for models that price by them; token pricing (gpt-image-1) is exact. Image input tokens of an edit are billed at the text input price. Migration ImageGeneration adds PerImageUsd. Before the call, the budget reserves the most the request could cost (n pictures, up to 6 400 output tokens each).

Limits and privacy. Edits may upload 50 MB in all (Gateway__MaxImageRequestBodyBytes, at most 16 pictures; raise the proxy's client_max_body_size with it). Pictures are held in memory for the duration of the request only, never written to disk or logged, and sent again from memory on fallback. A key whose Attached files setting is Text only cannot edit pictures (attachment_not_allowed); Images only may. The PII policy scans the prompt (and the other form fields) but cannot see inside pictures, and generated pictures are returned as the provider made them. Streaming (stream, partial_images) is not supported and answers 400 invalid_request. Pictures a provider returns as a URL (DALL·E) point at the provider's storage; ask for response_format: "b64_json" to keep them off third-party links.