Image generation - Triw

Triw relays image models through the OpenAI images API, so the same client code that works with OpenAI works here.

Generate an image

curl https://router.triw.io/v1/images/generations \
  -H "Authorization: Bearer $TRIW_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-1",
    "prompt": "A matte black mechanical keyboard on a walnut desk, soft window light, 50mm photo",
    "size": "1024x1024",
    "n": 1
  }'

Response follows OpenAI: base64 (default for gpt-image-*) or a hosted URL, depending on the model and the response_format you pass.

from openai import OpenAI
import base64, pathlib

client = OpenAI(api_key="sk-...", base_url="https://router.triw.io/v1")

r = client.images.generate(
    model="gpt-image-1",
    prompt="Isometric illustration of an API gateway, muted palette",
    size="1024x1024",
)
pathlib.Path("out.png").write_bytes(base64.b64decode(r.data[0].b64_json))

Edit an image

curl https://router.triw.io/v1/images/edits \
  -H "Authorization: Bearer $TRIW_API_KEY" \
  -F "model=gpt-image-1" \
  -F "image=@input.png" \
  -F "prompt=Replace the background with a plain studio grey"

/v1/images/variations is not implemented by the upstream providers Triw relays; expect a not-implemented response.

Available image models

Ask the gateway, not the docs — availability is per group:

curl https://router.triw.io/v1/models -H "Authorization: Bearer $TRIW_API_KEY" | grep -i image

Common identifiers include gpt-image-1, dall-e-3, dall-e-2, and flux-* families when those channels are connected.

Parameters that matter

Parameter	Notes
`size`	Must be supported by the model — `1024x1024`, `1024x1536`, `1536x1024` for `gpt-image-*`; `1024x1024`, `1792x1024`, `1024x1792` for `dall-e-3`
`n`	Some models only return 1 image per call
`quality`	`low` / `medium` / `high` for `gpt-image-*`; `standard` / `hd` for `dall-e-3`
`response_format`	`url` or `b64_json`; unsupported values are rejected upstream

Practical notes

Image calls are slow (10–60s). Set client timeouts above 120s; do not retry on timeout without de-duplicating, or you may pay twice.
Prompts are forwarded verbatim. Provider-side moderation applies — a rejected prompt returns invalid_request_error with the provider’s reason.
Billing is per image (plus tokens for gpt-image-* revisions), so check the price for the specific model before generating batches.