Images

Use your customer API Key at https://openai.2yanx.dpdns.org. Send Authorization: Bearer YOUR_API_KEY. Long requests should use SSE so the connection receives heartbeats while generation is pending.

Generate an image

curl -N https://openai.2yanx.dpdns.org/v1/images/generations \
  -H 'Authorization: Bearer YOUR_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{"model":"gpt-image-2","prompt":"A yellow circle on a white background","resolution":"4k","aspect_ratio":"16:9","stream":true,"delivery":"url"}'

With the site extension delivery:"url", the final event is image_url.completed. Download the image from the returned URL. This event requires custom client handling.

event: image_url.completed
data: {"type":"image_url.completed","operation":"generate","created_at":1790980000,"url":"https://s3.yanxinyu.ggff.net/media_outputs/0123456789abcdef0123456789abcdef.png","request_id":"REQUEST_ID"}

data: [DONE]

Omit delivery for the default SSE event: image_generation.completed, containing b64_json, actual size, output_format, created_at, url, and request_id. Even response_format:"url" retains Base64 in this default SSE mode. delivery:"url" requires stream:true and cannot be combined with response_format:"b64_json".

Models and fixed quality

Model	Fixed quality
gemini-3.1-flash-lite-image	standard; 1K only
gemini-3.1-flash-image	standard
gemini-3-pro-image	hd
gpt-image-2.5-fast	standard
gpt-image-2.5-pro	hd
gpt-image-2	hd
image-basic	standard
image-fast	standard
image-pro	hd
image-creative	hd

These are public product route names. Names do not guarantee an original-provider connection or a precise model snapshot. gpt-image-2.5-fast and gpt-image-2.5-pro are site product names, not verified official OpenAI model IDs.

Parameters and output dimensions

model and a non-blank prompt are required. Prompts accept at most 4096 Unicode characters. n is always 1; enhance defaults to true. quality may be omitted, auto, or the fixed value for that model. low, medium, and high are rejected.

resolution accepts 1k, 1080p, 2k, 4k, case-insensitively; default 1K. Lite rejects 2K/4K before submission. aspect_ratio defaults to 1:1 and accepts 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9.

size accepts auto or these tier/ratio presets:

Tier	Presets
1K	1024x1024, 1536x1024, 1024x1536, 1792x1024, 1024x1792
2K	2048x2048, 2048x1152, 1152x2048
4K	4096x4096, 4096x2304, 2304x4096

Explicit size, resolution, and aspect_ratio must agree. A preset selects a tier and ratio, not a guaranteed exact pixel resize. Recorded square 4K results: Flash/Pro returned 4096×4096; the three GPT Image routes and Basic returned 2880×2880. Recorded 4K 16:9 results for the GPT routes and Basic were 3840×2160. Basic square 1K/2K returned 1440×1440 / 1920×1920. Read dimensions from the actual output.

image_url accepts one HTTP(S) URL or Base64 image data URL; reference_images accepts a non-empty array of those sources. They are mutually exclusive. These, resolution, aspect_ratio, enhance, and delivery are site extensions, not universally supported original-provider fields.

Unknown fields are rejected. Masks, transparent-background control, output format/compression control, moderation, input fidelity, seed, style, and user fields are unavailable.

Edit with reference images

POST /v1/images/edits accepts JSON or multipart. JSON uses image, containing one source or an array:

{
  "model": "gemini-3-pro-image",
  "prompt": "Keep the composition and change red to blue",
  "image": ["https://example.com/reference.png"],
  "stream": true,
  "partial_images": 0
}

Multipart accepts image or repeated image[] files, at most 20 MiB per file. Boolean form fields are true/false. At least one reference is required; masks are unavailable. The default SSE completion event is image_edit.completed. With delivery:"url", the event is image_url.completed and operation is edit.

Non-streaming results and SDK

Without stream:true, omitted response_format returns an owned storage URL. Explicit b64_json adds Base64 to the successful JSON result. Decode the actual image format; output may be PNG or JPEG.

{"created":1790980000,"data":[{"url":"https://s3.yanxinyu.ggff.net/media_outputs/0123456789abcdef0123456789abcdef.png","b64_json":"BASE64_IMAGE"}]}

OpenAI Python SDK default SSE parsing works with site parameters in extra_body:

import base64
from pathlib import Path
from openai import OpenAI

client = OpenAI(base_url="https://openai.2yanx.dpdns.org/v1", api_key="YOUR_API_KEY", max_retries=0)
events = client.images.generate(
    model="gpt-image-2", prompt="A yellow circle on white",
    stream=True, partial_images=0,
    extra_body={"resolution": "4k", "aspect_ratio": "16:9"},
)
for event in events:
    if event.type == "image_generation.completed":
        Path("result." + event.output_format).write_bytes(base64.b64decode(event.b64_json))

Token usage is omitted because billing is per image. Strict consumers requiring original token usage must support its absence. quality:"auto" in the completion event does not assert an unverified original-provider quality tier.

Completion, errors, and capacity

SSE sends comment heartbeats while waiting, then a completion event or an error event. No intermediate previews are available; partial_images only accepts 0. HTTP 200 alone does not confirm generation success.

The entry admits up to 20 requests, including waiting and delivery; that does not mean 20 images generate concurrently. Excess requests receive 429 before generation submission. Non-streaming long requests can still encounter proxy timeouts.

A disconnected client does not cancel an already submitted generation. Do not automatically re-submit after a timeout or delivery error. Preserve X-Image-Request-Id or the final event's request_id, and check consumption with support first.