FLUX 3 Image alternatives on Sume, feature by feature

If you want what FLUX 3 Image does and have to stay on Sume, split it into features. Ten reference images: FLUX.2, Nano Banana, and most edit rows take up to 10, ChatGPT Image 2.5 takes 16. 4K output: Nano Banana 2 and Pro. Region editing: ChatGPT Image 2.5 with mask_url. Box-based layout, web grounding, and a seed have no Sume equivalent, because Sume's catalog does not list them for any image model.

This is a feature map, not a ranking. FLUX 3 Image launched on Oct 1 and I have no side-by-side output to compare, so nothing below says one model looks better than another.

What does FLUX 3 Image offer, per the vendor pages?

From Replicate, OpenRouter, BFL's overview, and press coverage (all read 2026-10-03): text-to-image and reference edits with up to 10 images; resolution tiers from 768sq to 4k; fixed aspect ratios from 21:9 to 9:21 plus auto; a grounding option, on by default, that searches the web or images before generating; a safety_tolerance range of 0 to 4; box-based scene composition on a 0-1000 grid; and reported multi-step edits that leave other parts of the image alone.

Open weights are reported as coming in a few weeks, with no date or license, and OpenRouter's listing allows one image per request.

Which Sume model covers which feature?

Every cell in the Sume column comes from Sume's image docs and catalog. A gap is stated as a gap.

Do ten references behave the same on every model?

A matching count is not matching behavior. The catalog says how many references a model will accept, not how many it will use well, and Sume's docs say nothing about attention across references. FLUX 3 Image's own format lets you tie a reference to a target box as ref_image_0 and so on; on Sume you tie a reference to a role in words, and with a mask edit the mask applies to the first image only, per OpenAI's guide. Put the image being edited first, then the style or object references, and number them in the prompt.

The practical test is cheap: send the same three-reference request to two models at low quality where the row allows it, and compare which reference each one ignored.

Which should you start with?

Pick by the job. For reference-heavy product scenes, start with FLUX.2 Pro, Nano Banana Pro, or ChatGPT Image 2.5, and use the model choice guide for photo edits, which lists reference limits per model. For a region change inside a finished image, use ChatGPT Image 2.5 and a mask, and, if the rest of the picture must not move, paste the original back after the call.

For a 4K deliverable, use Nano Banana Pro and read its endpoint record for price before you loop. For layout that has to land in exact positions, there is no Sume substitute; test FLUX 3 Image on a host that lists it. For ad text accuracy, the FLUX 3 or Ideogram 4.5 test plan lays out a fair comparison.

• Same job, many references: FLUX.2 Pro or Nano Banana Pro.
• Same job, one region: ChatGPT Image 2.5 with mask_url.
• Same job, large output: Nano Banana Pro at 4K.
• Let Sume choose: model: "sume/auto", which never discloses the model that ran.

Can Sume's Auto pick for you?

model: "sume/auto" lets Image Router choose the family. Sume's docs say it never discloses which one ran, it is not listed in GET /v1/images/models, and job.model stays sume/auto. That makes Auto convenient for everyday prompts and a poor choice when you must reproduce a result or attribute output to a specific model, which is the case if you are comparing against FLUX 3 Image.

Pin a model id for any comparison. The Image API choice guide covers the whole catalog.

How will you know when FLUX 3 arrives?

Watch GET /v1/images/models for a black-forest-labs/flux.3... id. When one appears, its supported_parameters settle every row of the table above, because Sume rejects any parameter a model does not list. Until then, the FLUX 3 versus FLUX 2 note is the one to read.