LLM Proxy

Appearance

Create image ​

Generate an image from a text prompt.

POST/v1/images/generations

JSON body. Unrecognized fields are forwarded to the provider as extension parameters rather than rejected — see Vendor extension fields.

Request
bash
curl http://localhost:8080/v1/images/generations \
  -H "Authorization: Bearer $LLM_PROXY_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-image-1", "prompt": "A lighthouse at dusk", "size": "1024x1024"}'
Request — OpenAI SDK
python
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="sk-your-proxy-key")
img = client.images.generate(model="gpt-image-1", prompt="A lighthouse at dusk")
print(img.data[0].url)
Response
json
{
  "created": 1718047200,
  "data": [
    {
      "url": "https://…/lighthouse.png",
      "revised_prompt": "A lighthouse at dusk, warm light"
    }
  ],
  "usage": {
    "input_tokens": 50,
    "output_tokens": 1240,
    "input_tokens_details": {"text_tokens": 50, "image_tokens": 0}
  }
}

Request fields ​

prompt (required), model, n (1–12), quality, size, style, response_format (url default, or b64_json), user, background, moderation, output_format, output_compression, partial_images, stream.

  • size: auto, or WIDTHxHEIGHT where both edges ≥ 16 and divisible by 16, the larger edge ≤ 3840, and the aspect ratio within 1:3–3:1.
  • n: the schema accepts 1–12 (the widest documented provider range). Providers apply their own caps; see Vendor extension fields.
  • gpt-image-* models additionally receive background, moderation, output_compression, output_format, partial_images; response_format and style are omitted for them.

Vendor extension fields ​

Fields outside the OpenAI Images schema are passed through to the provider's own request body instead of being rejected, so upstream-specific options stay reachable without extra_body (which the OpenAI SDK offers for chat, not images).

qwen / qwen-intl (DashScope) accept:

  • image — a public image URL, a base64 data URL, or an array of either. Supplying it turns the call into image-to-image; omitting it is text-to-image.
  • negative_prompt, seed, prompt_extend, prompt_extend_mode, enable_thinking, watermark — forwarded into DashScope's parameters block.
  • On /v1/images/edits: color_palette, bbox_list, enable_sequential, thinking_mode.

qwen/qwen-intl per-model caps (violations return 400 invalid_request_error): n is 1–6 for qwen-image-* and 1–4 for wan* (1–12 with enable_sequential).

Strictness is configurable: with server_params.unknown_fields_policy: error the proxy rejects unknown fields instead of forwarding them (per-provider EXEMPT_EXTRA_KEYS still keeps documented upstream fields working).

Response ​

{"created": <int>, "data":[{"url"|"b64_json", "revised_prompt"?}], …} with optional usage (input_tokens, output_tokens, input_tokens_details.{text_tokens,image_tokens}).

  • There is no model field in image responses, and the proxy does not convert between url and b64_json — you get whatever the provider produced.
  • stream: true emits SSE events image_generation.partial_image (with b64_json, partial_image_index) and image_generation.completed, then [DONE].
  • There is no async/background image mode.

Provider support ​

openai, the openai-compatible family, gemini, and qwen providers can serve image requests — see the capability matrix and Vendor extension fields.

Request
bash
curl http://localhost:8080/v1/images/generations \
  -H "Authorization: Bearer $LLM_PROXY_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-image-1", "prompt": "A lighthouse at dusk", "size": "1024x1024"}'
Request — OpenAI SDK
python
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="sk-your-proxy-key")
img = client.images.generate(model="gpt-image-1", prompt="A lighthouse at dusk")
print(img.data[0].url)
Response
json
{
  "created": 1718047200,
  "data": [
    {
      "url": "https://…/lighthouse.png",
      "revised_prompt": "A lighthouse at dusk, warm light"
    }
  ],
  "usage": {
    "input_tokens": 50,
    "output_tokens": 1240,
    "input_tokens_details": {"text_tokens": 50, "image_tokens": 0}
  }
}

MIT licensed

LLM Proxy