FLUX.2 pro prompt upsampling is on by default; Sume has no switch

On FLUX.2 pro, the vendor describes automatic prompt upsampling as enabled by default, and Sume gives you no field to turn it off: its Image API docs say allowed_passthrough_parameters is empty for every endpoint in v1 and provider.options must be omitted or empty. If your prompt contains words that must appear exactly, plan for the model to interpret your prompt rather than copy it, and test with several candidates.

I cannot tell you from the sources how often upsampling changes quoted text. What the sources do establish is who controls the setting: the vendor default applies, and a Sume request does not override it.

What the fal page says

The fal page for FLUX.2 [pro] describes a streamlined pipeline with zero-configuration quality, meaning no inference step or guidance parameter to tune, and lists automatic prompt upsampling as enabled by default. It also says the model supports JSON structured prompts and HEX color code control, and multi-image referencing with an @ syntax. The price is $0.03 for the first megapixel of output plus $0.015 per extra megapixel of input and output, with the page's examples at 1024x1024 for $0.03 and 1920x1080 for $0.045.

Because those prompt features are in the prompt text, they pass through a Sume request unchanged: write a JSON-style prompt or a HEX value in the prompt string like any other text.

What a Sume request can and cannot set

In the request builder on origin/main, a FLUX.2 pro call carries the prompt, any reference image URLs, the image size, the image count, the output format and a safety-checker flag. There is no upsampling field in that payload. The docs back this up from the public side: unlisted parameters return 400 unsupported_parameter, and passthrough options are empty.

A test that fits an exact-text prompt

Because you cannot pin the behavior, measure it. Generate four candidates of the same prompt with n: 4, look at the quoted words in each, and decide whether the model meets your bar. The request below asks for a 4:5 poster; the response is 200 with data[].url for each image, or 202 with a job to poll if the generation outlasts the 30-second wait.

If the words are the whole point, run the same prompt through a model built around text, and compare. Sume's catalog includes Ideogram 4.5, and the fal page for it advertises accurate text rendering, so it is the obvious control.

Keep the prompt itself short and literal when exact words matter: put the string in quotes, say where it sits in the frame, and avoid adjectives the model can reinterpret. After the run, compare the four candidates with the prompt side by side and keep notes on how many reproduced the string exactly, so the next decision rests on your own count rather than on a vendor claim.

import os
import requests

H = {"Authorization": "Bearer " + os.environ["SUME_API_KEY"]}
body = {
    "model": "black-forest-labs/flux.2-pro",
    "prompt": 'Poster with the exact words "OPEN LATE" in thick white letters',
    "aspect_ratio": "4:5",
    "n": 4,
}
r = requests.post("https://api.sume.com/v1/images", headers=H, json=body, timeout=60)
print(r.status_code)
if r.status_code == 200:
    for image in r.json()["data"]:
        print(image["url"])