Evolink Image

Evolink GPT Image 2

Async image generation via Evolink's gpt-image-2 endpoint. Supports text→image, image→image, and edit-with-reference. All work goes through the bundled script — do not hand-roll HTTP calls unless the user asks for raw API use.

When to reach for this skill

Trigger on requests to produce a bitmap image from a description: "generate", "create an image", "draw", "render", "make a picture of", "edit this photo", "change the background of this image", "turn this sketch into …". Applies equally when the user just drops a prompt ("A cinematic sunset over Shenzhen skyline") without the word "image" — the intent is clear.

Skip when the user wants vector/UI/code-driven art — frontend-design, algorithmic-art, shader-dev, or canvas-design handle those better.

Environment

The script reads EVOLINK_API_KEY from the environment. It is already configured globally; no action needed on a fresh run. If the call returns 401, tell the user their key is missing or expired rather than trying to patch it yourself.

The script

scripts/generate_image.py --prompt "..." [options]

Full flag reference (all optional except --prompt):

The script blocks until the task finishes (or fails / times out), downloads each result to <output-dir>/<task-id>-<index>.<ext>, and prints the saved paths. Result URLs expire in 24 h, so always download unless the user explicitly just wants the URL.

Common invocations

Plain text-to-image

uv run scripts/generate_image.py -p "A beautiful colorful sunset over the ocean"

Widescreen, high quality

uv run scripts/generate_image.py \
  -p "Cinematic wide shot of a futuristic city skyline at dusk" \
  -s 16:9 -r 4K -q high

Image-to-image edit

uv run scripts/generate_image.py \
  -p "Add a cute cat next to her" \
  -i https://example.com/input.png \
  -s 1:1 -q medium

Batch of 4

uv run scripts/generate_image.py \
  -p "A cute robot in pixel art style" \
  -s 1:1 -r 2K -q high -n 4

Just want the URL (no download)

uv run scripts/generate_image.py -p "..." --no-download

Behaviour notes

• Async under the hood. The API returns a task-unified-… id; the script polls /v1/tasks/{id} every few seconds until status is completed or failed. The user sees only the final result.
• Size vs resolution. resolution (1K/2K/4K) is only a hint when size is a ratio. If the user gives explicit pixels (1024x1024), drop --resolution. Each side must be a multiple of 16; aspect ≤ 3:1; total pixels 0.65 MP – 8.29 MP.
• Cost awareness. quality=high is roughly 4× the token cost of medium, and n=4 quadruples again. Don't bump both without a reason — default to medium + n=1 unless the user asks otherwise.
• Failures. On status=failed the task response carries error.code / error.message; surface those verbatim. Common ones: content_policy_violation (rewrite prompt), insufficient_quota (402), rate_limit_exceeded (429, retry later).
• Reference image input. The URLs must be publicly reachable by Evolink's servers and end in an image extension, ≤ 50 MB each, .jpg/.png/.webp. If the user gives you a local file, upload it somewhere first (or tell them you need a URL) — don't try to base64 it; the API doesn't accept that here.

Reporting back to the user

After a successful run, state the saved path(s) and, if relevant, the task id. Example: "Saved to evolink-out/task-unified-1757156493-imcg5zqt-0.png." Don't re-paste the full URL unless they asked for it — the file is what matters.