AI action figure box from a selfie: a gpt-image-2.5 reference edit

To make the action-figure-in-a-box image from a selfie, call POST /v1/images with model: "openai/gpt-image-2.5", put the selfie in input_references, and describe the toy, the blister pack and the card in the prompt. Ask for a blank name panel and add the name yourself afterwards, because exact lettering is the part of a packaging image most worth proofreading.

The selfie has to be a public HTTPS image URL. Sume's Image API docs reject localhost, private-network and non-HTTPS references before submission, so host the photo somewhere fetchable first.

The request

gpt-image-2.5 takes up to 16 references, so the selfie can travel with a second image per accessory (a mug, a camera, a pet). Use aspect_ratio: "auto" on edits so the output follows the reference; the docs state that omitting the field is not the same as auto. Ask for n: 4 and pick the best likeness; read the n range descriptor in the catalog first, since per-model ceilings sit below the 10-per-call maximum.

{
  "model": "openai/gpt-image-2.5",
  "prompt": "A collectible action figure of the person in the reference photo, sealed in a retro blister pack on a printed card. Keep the face, hair and glasses from the reference. Leave the name panel at the top of the card blank. Studio product photo, soft light.",
  "input_references": [
    {"type": "image_url", "image_url": {"url": "https://example.com/selfie.jpg"}}
  ],
  "aspect_ratio": "auto",
  "quality": "high",
  "n": 4
}

Put the name on the card in code

A name, a tagline and a 'Series 1' line are exact strings. Drawing them in Pillow or in HTML over the generated image means the spelling is whatever you typed. If you do ask the model to render them, put the text in quotes in the prompt and read the result letter by letter before you share it.

import os
import requests


def generate(payload):
    response = requests.post(
        "https://api.sume.com/v1/images",
        headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
        json=payload,
        timeout=60,
    )
    response.raise_for_status()
    return response.status_code, response.json()

# status 200: body["data"][i]["url"] are the four images
# status 202: the 30-second wait ran out; poll the job instead
status, body = generate({
    "model": "openai/gpt-image-2.5",
    "prompt": "Collectible action figure in a blister pack, blank name panel",
    "input_references": [{"type": "image_url", "image_url": {"url": "https://example.com/selfie.jpg"}}],
    "aspect_ratio": "auto",
    "n": 4,
})
print(status)

Fields this recipe uses

Cost and rules

Sume's docs give the gpt-image-2.5 output estimate at 1024x1024 as $0.09366 for xhigh and $0.21072 for max before input tokens and Sume pricing. Read the cost_usd pricing line from the model's endpoint record for what you will actually pay, and multiply by n. A generation that fails is not billed.

Use photos of yourself or of people who agreed to appear in the image, and keep the final wording of the card (series name, any trademarked franchise) your own.