FLUX 3 Image grounding is on by default: web search and Sume

In FLUX 3 Image, grounding is a boolean that defaults to true: the model can research your prompt with web search and image search before it generates, and false turns both off. Sume's image request table has no equivalent field, and a field a model does not list is rejected with 400 unsupported_parameter, so on Sume you supply the visual target yourself with references.

BFL facts are from its FLUX 3 Image reference; Sume facts from the Image API and catalog code, read 2026-10-01.

What does grounding change?

With the default on, a prompt naming a real product, place or event can be informed by search results before generation. Two things follow for an integration. First, the same prompt may not reproduce exactly, because what the search finds is outside your request. Second, if you need outputs that depend only on your prompt and references, send grounding: false. The reference does not say what the model searches for or how results are used beyond that sentence.

What does Sume do with an unknown field?

The docs say a request that sets a parameter the selected model does not list is rejected with 400 unsupported_parameter rather than silently dropped. So sending grounding to POST /v1/images would fail instead of being ignored. The Sume catalog for FLUX lists flux-2-pro and flux-2-flex in code, not FLUX 3 Image.

How do I give Sume a visual target?

Send the images you want matched as input_references, which the docs describe as reference images for image-to-image, with public HTTPS URLs. Pick a model whose catalog marks it edit-capable; text-only models reject references. Describe in the prompt how each reference should be used. For the closest comparison on an earlier model, read FLUX.2 Max grounding and what Sume lists.

Should I turn grounding off on BFL?

Turn it off when repeatability matters, such as a brand template, or when the prompt must not pull outside imagery. Leave it on when you want the model to look up an unfamiliar subject. Test both on a handful of prompts before choosing; the reference states no accuracy numbers and this post makes no claim about which output looks better.