Character turnaround sheet by API: front, profile, back views

To get a front, profile and back view of one character by API, render the front view first, then make one request per extra angle and send the front image as a reference with a prompt such as "the same woman, in profile looking right". Google's Gemini image docs describe this as the way to build a 360-degree view: prompt for each angle in turn and include previously generated images so the character stays consistent. On Sume that is POST /v1/images with input_references.

Google's method comes from Image generation with Gemini; Sume's side comes from the Sume Image API page. Both were read on 2026-10-02.

What does Google say about building a 360 view?

The Gemini page lists "Character consistency: 360 view" as a prompt pattern. Its text says you can generate 360-degree views by iteratively prompting for different angles, that you should include previously generated images in later prompts to maintain consistency, and that a complex pose needs a reference image of that pose.

Its template is A studio portrait of [person] against [background], [looking forward/in profile looking right/etc.]. The page's own example sends one photo and asks for the same man against white, in profile looking right. Nothing in it promises an exact match; it describes a method, not a guarantee.

How do I do the same on Sume?

Send the first render back through input_references. The Sume docs say reference URLs must be public HTTPS, and that on edit and image-to-image calls you should prefer aspect_ratio: "auto" to match the reference, because omitting the field is not the same as auto.

The script below makes the front view, then asks for profile and back with that image as the only reference. It stops on 202, because a request that outlasts the 30-second blocking budget returns a job envelope, not an image.

import os, requests

URL = "https://api.sume.com/v1/images"
HEAD = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
MODEL = "google/nano-banana-2"

def make(prompt, ref=None):
    body = {"model": MODEL, "prompt": prompt}
    if ref:
        body["aspect_ratio"] = "auto"
        body["input_references"] = [
            {"type": "image_url", "image_url": {"url": ref}}
        ]
    r = requests.post(URL, headers=HEAD, json=body, timeout=60)
    if r.status_code == 202:
        raise SystemExit("202: read the job result instead")
    r.raise_for_status()
    return r.json()["data"][0]["url"]

front = make("Studio portrait of a woman with a green rain jacket, "
             "looking forward, plain white background")
for view in ("in profile looking right", "from behind"):
    prompt = ("The same woman, same jacket and hair, "
              f"studio portrait on white, {view}")
    print(view, make(prompt, front))

How many references can one request carry?

Read the input_references range from the catalog before you plan a sheet with a pose guide plus the front view. The docs page states that models whose range is {"min": 0, "max": 0} are text-to-image only and reject references.

What if Sume cannot fetch the reference image?

Sume's errors page lists image_not_fetchable and input_media_unreachable for media it cannot fetch or mirror safely, and says to check that the input is a public HTTPS image URL. If the URL from a previous result is not accepted as a reference in your setup, upload the image to a public HTTPS location you control and send that URL instead.

Keep the wording of the character description identical across the three prompts. Only the angle phrase should change; that is the one variable you want to test.