gpt-image-2 — Image Generation & Editing

OpenAI gpt-image-2 through AceDataCloud. Two endpoints, both synchronous (return image url(s) directly). Its standout is editing: feed real images (logos, QR codes, product shots, screenshots) and it composites/restyles them faithfully — great for on-brand video assets and character consistency.

Setup: See authentication for token setup.

1. Generate (text → image)

curl -X POST https://api.acedata.cloud/openai/images/generations \
  -H "Authorization: Bearer $ACEDATACLOUD_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-image-2","prompt":"a clean dark tech hero background with a glowing API hub, lots of negative space","size":"1792x1024","n":1}'

2. Edit / composite (images + prompt → image) ← the powerful one

Multipart. Pass one or more source images via repeated image[] (local files with @, or URLs). Use it to fuse a real logo/QR into a generated scene, keep a subject consistent across scenes, or restyle a screenshot.

curl -X POST https://api.acedata.cloud/openai/images/edits \
  -H "Authorization: Bearer $ACEDATACLOUD_API_TOKEN" \
  -F "model=gpt-image-2" \
  -F "prompt=Place this logo crisply in the top-left on the tech background; keep the logo's exact colors and shape." \
  -F "image[]=@background.png" \
  -F "image[]=@logo.png" \
  -F "size=1792x1024" \
  -F "n=1"

Response (both endpoints): {"data":[{"url":"https://...png"}]} → download data[0].url.

Sizes

size is WxH (a preset) or "auto". Common presets:

Aspect	Sizes
16:9	`1792x1024` (HD), `2048x1152`, `3840x2160` (4K)
9:16	`1024x1792`, `1152x2048`, `2160x3840`
1:1	`1024x1024`, `2048x2048`, `2880x2880` (4K)

(Omit size or use "auto" to let the model pick. Custom sizes must have both sides a multiple of 16, each side ≤ 3840, total pixels between 655,360 and 8,294,400, and an aspect ratio ≤ 3:1 — otherwise 400.)

Tips

Editing keeps things faithful — to place a logo/QR exactly, pass it as one of the image[] and say "keep its exact colors/shape, do not redraw it".
For character/scene consistency across video beats, generate one hero image, then edits it per beat instead of regenerating from scratch.
Text in images renders legibly — good for titles/labels you don't want to overlay in HTML.
n accepts 1–10 and you are billed per image returned, not per request — n: 4 costs 4×. (response_format: "b64_json" still requires n: 1.)
Both endpoints are synchronous by default. For long 4K jobs pass callback_url (optionally with async: true) and poll POST /openai/tasks with {"id": "<task_id>"}.