FLUX 3 Image vs Nano Banana Pro for 4K editing on Sume

If you need 4K editing from a reference image on Sume today, use Nano Banana Pro: its catalog row publishes a resolution of 512, 1K, 2K, or 4K and takes up to 10 reference images. FLUX 3 Image, which launched on Oct 1 with 4K output, 10 references, and box-based region editing, is not in Sume's catalog, so the comparison is between a model you can call and one you can only read about.

This post lays the two side by side using vendor pages for FLUX 3 Image and Sume's own catalog rules for Nano Banana Pro. It makes no quality claim, because there is no side-by-side test of the two on the same inputs here.

What does each model claim for editing?

For FLUX 3 Image, the claims are the vendor's and press coverage of them. Replicate lists up to 10 reference images, a resolution of 768sq, 1k, 1.5k, 2k, or 4k, and an aspect ratio from 21:9 down to 9:21 (read 2026-10-03). The Decoder reports multi-step edits that leave other parts of the picture alone and scene composition with bounding boxes. Tech Times reports a native output of about 5,456 by 3,072 pixels.

For Nano Banana Pro on Sume, the claims are descriptors you can query: the catalog row for the model lists input_references from 0 to 10, a resolution enum, an aspect_ratio list, and n up to 4. Sume's pricing notes say the 4K list price is higher than the default tier and that admission scales for 4K.

How do they line up, feature by feature?

The table uses only statements from the sources named in its caption.

How are reference images supplied?

Tech Times reports that FLUX 3 Image takes references as URLs or base64 data and cites them in the prompt as ref_image_0 through ref_image_9; BFL's overview shows the same tokens inside its box rows. Sume takes references only as public HTTPS URLs in input_references, and rejects localhost, private-network, and non-HTTPS URLs before the job starts. There is no token syntax for pointing at a reference: you refer to them in plain words, and for several references, numbering them in the prompt ("the first image is the room, the second is the sofa") is the safe habit.

If your pipeline keeps images in private storage, the Sume route needs a signed or public HTTPS link generated per request. That is an extra step you would not have with base64 on a FLUX 3 host, and it is the first thing to budget when you compare the two.

What can you do on Sume that approximates box placement?

Nothing in Sume's catalog takes coordinates. Two workarounds exist. The first is to describe the position in the prompt ("the mug in the lower right third, cropped by the frame") and send a reference image of the mug; the second is a mask edit on ChatGPT Image 2.5, where you paint the target region. The bounding box versus mask note walks through the second one.

For Nano Banana Pro itself, prefer reference images over long descriptions when identity matters; the Nano Banana reference limits post covers how many of each kind Sume accepts.

A 4K edit request on Sume

Set resolution to 4K and aspect_ratio to auto so the edit follows the reference. Sume's docs say that on edit calls auto matches the reference, and that omitting the field is not the same as auto. Because 4K is one of the slow configurations, handle the 202: the script prints either the image URL or the status URL to poll. The reference must be a public HTTPS URL.

import os
import requests

r = requests.post(
    "https://api.sume.com/v1/images",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    json={
        "model": "google/nano-banana-pro",
        "prompt": "same room, repaint the walls sage green, keep furniture",
        "resolution": "4K",
        "aspect_ratio": "auto",
        "input_references": [
            {"type": "image_url", "image_url": {"url": "https://example.com/room.jpg"}}
        ],
    },
    timeout=60,
)
body = r.json()
print(body["data"][0]["url"] if r.status_code == 200 else body)

How should you decide?

Choose on what you need to ship this week. If the job needs exact box placement and you can use a hosted API outside Sume, FLUX 3 Image is the model to test; BFL's overview documents the box format and Replicate hosts it. If the job needs a 4K edit inside one account with the rest of your Sume work, Nano Banana Pro is the callable choice.

Run the same three reference-edit prompts on both before you pick, and compare the files, not the launch posts. The FLUX.2 versus Nano Banana Pro specs post is the closest Sume-side data point, and the 4K wait note explains why a 4K call may return a job.