FLUX 3 Image on Replicate: safety_tolerance 0-4 vs Sume

Replicate's FLUX 3 Image page lists eight inputs: prompt, images (up to 10), aspect_ratio, resolution, grounding, safety_tolerance (0 strictest to 4 most permissive), output_format, and output_quality. Of those, Sume's image API has direct equivalents for the prompt, the reference images, the aspect ratio on most models, and the output format. It has no safety dial, no grounding switch, and no lossy-quality slider, and it does not list FLUX 3 Image yet. Sending an unlisted field returns 400 unsupported_parameter, not a silent drop.

This post is a field map for people reading the Replicate form who need to know what survives a move to Sume's FLUX.2 rows, the nearest Black Forest Labs models in the catalog.

What does Replicate's FLUX 3 Image form list?

From the model page, read 2026-10-03: aspect_ratio is auto by default or a fixed ratio from 21:9 down to 9:21; resolution is 768sq, 1k (default), 1.5k, 2k, or 4k; grounding is on by default and runs a web or image search before generation; safety_tolerance runs 0 to 4; output_format is webp, jpg, or png. Reference images must be at least 256 by 256 pixels and at most 16 megapixels, in jpeg, png, gif, or webp.

The page gives no pricing, license, or bounding-box format, so none of that is covered here; Black Forest Labs' own docs describe the box format.

Which of those fields map to Sume?

Sume's published parameters for FLUX.2 Pro and Flex come from the model catalog: prompt, n, aspect_ratio, input_references, and output_format. The mapping below uses the catalog and Sume's request-parameter table.

What do I do about the safety and grounding settings?

Nothing, because Sume exposes no dial for either. There is no safety_tolerance to raise, so a prompt that a permissive Replicate setting lets through may be refused on Sume with an error on the job, and Sume's docs say a failed generation is not billed. Treat the refusal as a result to handle rather than a setting to tune.

Grounding is the reverse problem: on Replicate it is on by default, which means the same prompt can produce different images depending on what the search finds. On Sume, a FLUX.2 prompt is read as written, so if a prompt depends on a real product or a current event, put the facts in the prompt or send a reference image of the subject.

Why the ratio and format words trip up ports

Two small spellings cause most 400s. Sume wants jpeg where Replicate says jpg, and Sume's FLUX.2 ratio list does not contain auto, so a script that defaults to auto needs a real ratio chosen per request. The cleanest way to find out what a model takes is to ask the catalog, as in the call below, which prints every parameter name FLUX.2 Pro advertises.

import os
import requests

r = requests.get(
    "https://api.sume.com/v1/images/models",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    timeout=30,
)
r.raise_for_status()
for m in r.json()["data"]:
    if m["id"] == "black-forest-labs/flux.2-pro":
        for name, d in m["supported_parameters"].items():
            print(name, d)

How should I handle resolution without a tier?

Replicate's 768sq to 4k tiers have no FLUX.2 counterpart on Sume, because the catalog publishes no resolution descriptor for those rows. Sume's docs say the FLUX family takes custom pixels through image_size on the legacy Image 1.0 route and that a 4:5 ratio is Instagram portrait, but they do not publish an output pixel count for FLUX.2 requests, so a post that quotes one would be guessing.

The practical approach is to generate at the ratio you want, read the width and height of the file that comes back, and, if you need more pixels, add a separate step. Sume has an upscaler at POST /v1/image-upscale-1.0/upscale, covered in the edit first, then upscale note. If you need a documented 2K or 4K tier on a generation call, Nano Banana 2 and Pro publish one, as the FLUX.2 versus Nano Banana Pro specs show.

What stays the same when you move?

Reference rules are close. Replicate wants images of at least 256 by 256 pixels; Sume wants public HTTPS URLs, and rejects localhost, private-network, and non-HTTPS URLs before submission. If your Replicate script passes file uploads, host the files first. The reference URL checklist lists the rejections.

Results differ in kind: Sume returns a hosted URL in data[].url, and slow generations come back as a 202 job after 30 seconds, as the Image API docs describe. The unknown-field note compares the error shapes, and the grounding note covers the search default in more depth.