Midjourney V8.2 personalization profiles: the Sume substitute

Midjourney V8.2 ships better personalization: profiles built from the images you have rated, with a larger pool of images to learn from. The Sume image API has no per-user taste profile and does not learn from your ratings. What you can do is write your taste down: a short brand brief in the prompt and a fixed set of reference images in input_references. That gives consistent output you can version and share.

What Midjourney says

From the Version 8.2 post and the 9/23 changelog.

What replaces a profile on Sume

A profile is hidden state. On Sume you make that state explicit and put it in the request. Two parts carry most of the effect.

• A brief of three to five sentences: subject matter, lighting, palette, mood, and what to avoid, written positively.
• Reference images chosen from your own best work, sent as input_references. Models differ in how many they take, so read the input_references range for the model you pinned.
• A pinned model id and quality, so a change of model never silently changes the look.

A reusable profile file

Store the profile as data so it is reviewed like code. The request builder reads it and sends the brief as the start of every prompt.

import json, os, requests

profile = {
    "model": "openai/gpt-image-2.5",
    "brief": "Warm natural light, muted earth palette, calm minimal sets, no neon.",
    "refs": ["https://example.com/ref1.jpg", "https://example.com/ref2.jpg"],
}

def render(subject):
    r = requests.post(
        "https://api.sume.com/v1/images",
        headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
        timeout=60,
        json={
            "model": profile["model"],
            "prompt": f"{subject}. {profile['brief']}",
            "aspect_ratio": "auto",
            "input_references": [
                {"type": "image_url", "image_url": {"url": u}} for u in profile["refs"]
            ],
        },
    )
    return r.status_code, r.json()

What you give up

A rating-based profile improves as you rate more images. A file does not, unless you update it. Review it monthly: add new best-of references, drop stale ones. Also note that references steer a model but do not train it; Sume's image API lists no fine-tuning or custom-model endpoint in the docs.

See the related post on style references for how the reference approach compares with Midjourney's sref codes.

Checks before you rely on it

Run the profile through a few unrelated subjects and look for the same palette and mood. If a reference dominates too much, swap it for a plainer one. If the brief is ignored, shorten it.