Image generation

Access is enabled separately

These instructions are available to every customer, but image generation is enabled only under a separate agreement with support@omev.be. A text API key or access to this page does not mean image products are active for the key.

Maximum reference images per request: 9. Reference (image-to-image) inputs are supported by image jobs with input_urls.

All image products generate 1K-class output (up to a 1024-pixel edge for square requests, or 1536x1024 / 1024x1536 where the shape offers it). There are no 2K or 4K tiers; asking for one answers 400.

There are two ways to make images, built for two different workloads.

Image jobs — the main path for pipelines. You place an order, poll for the result, and download the finished image from a URL we host for 48 hours. One model, omev-image-jobs, 1K-class output. This is the path for unattended bulk generation: your worker queues orders and picks up results, no human waiting on a socket.

Direct models. The image comes back inside the response body of a single call: omev-image-lite, omev-image-main and omev-image-pro. Choose these when someone is waiting for the picture on screen right now.

Image jobs API

Three calls, same host and same key as everything else.

1. Place an order

curl -sS https://llmapi.omev.be/api/v1/jobs/createTask \
  -H "Authorization: Bearer $OMEV_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "omev-image-jobs",
    "input": {
      "prompt": "Flat vector infographic: three steps of onboarding, numbered 1 2 3, short labels",
      "aspect_ratio": "1:1"
    }
  }'

{"code": 200, "msg": "success", "data": {"taskId": "task_omevimage_1754937600123_a1b2c3d4"}}

Field	Required	Values
model	yes	omev-image-jobs
input.prompt	yes	free text; spell out every word that must appear in the picture
input.aspect_ratio	no	auto (default, square), 1:1, 3:2, 2:3
input.resolution	no	1K — the only tier offered; anything else answers 400
input.quality	no	low (default) or auto; medium/high answer 400
input.input_urls	no	1–9 public http(s) reference image URLs (see below)
callBackUrl	no	we POST the final task record there; plain polling works too

A 403 normally means the key is not enabled for image jobs. Contact support@omev.be to agree access; do not repeatedly retry. A 400 answers {"code": 400, "msg": "<what was wrong>"}.

2. Check the result

curl -sS "https://llmapi.omev.be/api/v1/jobs/recordInfo?taskId=task_omevimage_1754937600123_a1b2c3d4" \
  -H "Authorization: Bearer $OMEV_API_KEY"

{
  "code": 200,
  "msg": "success",
  "data": {
    "taskId": "task_omevimage_1754937600123_a1b2c3d4",
    "model": "omev-image-jobs",
    "state": "success",
    "resultJson": "{\"resultUrls\":[\"https://llmapi.omev.be/api/v1/jobs/result?id=task_omevimage_1754937600123_a1b2c3d4\"]}",
    "failCode": "",
    "failMsg": "",
    "costTime": 96000,
    "createTime": 1754937600123,
    "completeTime": 1754937696123,
    "progress": 100
  }
}

state walks waiting → queuing → generating → success (or fail, with failCode and failMsg telling you why). Poll every 2–5 seconds and back off; a typical job takes one to five minutes. resultJson is a JSON string: parse it and take the URLs from resultUrls.

3. Download the image

Result URLs are public — no key is needed to fetch them — and stay valid for 48 hours. Pull the bytes into your own storage right away; after 48 hours the file is removed and the URL answers 404.

If your code already drives a task-based image API of this shape (create a task, poll a record, read resultUrls), it works here after two edits: the base URL and the key.

Reference images (image-to-image)

Orders can carry 1–9 reference images: add an input_urls array to input with public http(s) URLs of the pictures to edit or combine. The order of the list is preserved: if your prompt says "put the product from image 1 into the scene from image 2", image 1 is input_urls[0].

{
  "model": "omev-image-jobs",
  "input": {
    "prompt": "Put the product from image 1 into the environment from image 2",
    "input_urls": [
      "https://example.com/product.png",
      "https://example.com/location.jpg"
    ],
    "aspect_ratio": "1:1"
  }
}

• Omitting input_urls (or an empty list) keeps the plain text-to-image order.
• More than 9 URLs are rejected with a deterministic 400: A maximum of 9 reference images is allowed per request.
• References are copied to internal storage when the order is placed, so a source URL that expires later cannot break a running job.

The alias gpt-image-2-image-to-image selects the same pipeline, for integrations that already use that name.

omev-image-main

The request shape is the OpenAI-compatible Images API, so any SDK built for it works unchanged: point base_url at Omev and call images.generate. omev-image-main is the steady default for interactive use.

curl -sS https://llmapi.omev.be/v1/images/generations \
  -H "Authorization: Bearer $OMEV_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "omev-image-main",
    "prompt": "A photorealistic cat wearing sunglasses",
    "size": "1024x1024"
  }'

Field	Required	Values
prompt	yes	free text
size	no	1024x1024 (default), 1536x1024, 1024x1536, auto
quality	no	low (default) or auto; medium/high answer 400
n	no	1 only
response_format	no	b64_json only

The image sits in data[0].b64_json. If the call times out with 504, send the same request again with the same Idempotency-Key header: the order continues server-side and the retry picks up the finished result instead of paying for a second image.

omev-image-pro

Same OpenAI-compatible Images API shape as omev-image-main. omev-image-pro is the quality lane: prompts that carry embedded text (infographics, diagrams, charts, tables, labels, logos) render reliably here.

curl -sS https://llmapi.omev.be/v1/images/generations \
  -H "Authorization: Bearer $OMEV_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "omev-image-pro",
    "prompt": "Flat vector infographic: three steps of onboarding, numbered 1 2 3, short labels",
    "size": "1024x1024"
  }'

Field	Required	Values
prompt	yes	free text; spell out every word that must appear in the picture
size	no	1024x1024 (default), auto, or any WxH up to 1024 on each edge
quality	no	low, medium, auto; high answers 400
n	no	up to 4 images
response_format	no	b64_json only
background	no	transparent, opaque, auto
output_format	no	png, jpeg, webp

client = OpenAI(base_url="https://llmapi.omev.be/v1", api_key=os.environ["OMEV_API_KEY"])

resp = client.images.generate(
    model="omev-image-pro",
    prompt="Flat vector infographic: three steps of onboarding, numbered 1 2 3",
    size="1024x1024",
)
png_bytes = base64.b64decode(resp.data[0].b64_json)

The image sits in data[0].b64_json.

omev-image-lite

A different route and a different body shape — the one built around contents and parts. If your code already speaks that shape, it works here without changes. Output is always 1024×1024, so there is no size parameter.

curl -sS https://llmapi.omev.be/v1beta/models/omev-image-lite:generateContent \
  -H "Authorization: Bearer $OMEV_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{"parts": [{"text": "A photorealistic cat wearing sunglasses"}]}]
  }'

Response:

{
  "candidates": [{
    "content": {
      "role": "model",
      "parts": [{"inlineData": {"mimeType": "image/png", "data": "<base64 PNG>"}}]
    },
    "finishReason": "STOP"
  }]
}

The image sits in candidates[0].content.parts[0].inlineData.data, with its type in the neighbouring mimeType.

Set generationConfig.candidateCount (1–4) to receive that many separate candidates, each carrying one image in inlineData.

Errors

Same shape as the rest of the API — see Errors. A malformed body gives 400 with what was wrong; a refused prompt gives 400 as well, not a blank image. Image jobs report failures in the task record instead: state: "fail" with failCode/failMsg.