FLUX Kontext Steps: Local Dev Settings vs. Hosted Pro

You can set FLUX Kontext sampling steps in a local Kontext [dev] pipeline. In the Kontext editor on flux-3image, hosted [pro] and [max] have no step-count control. Each Generate click submits one editing request; the sampling iterations inside that request are handled by the provider.

Which kind of Kontext steps do you mean?

Sampling steps are the iterations used to render one result. Successive edits are separate requests: change the background, then upload that result to change the jacket. Those are two edits, each with its own internal sampling process.

As of October 2026, fal's hosted Kontext [pro] input schema lists a prompt, one image URL, seed, aspect ratio, and guidance, but no sampling-step field. Guidance is a separate setting, which flux-3image also does not expose. Typing “use 30 steps” into the prompt does not set a sampling parameter.

How many steps should you start with in local Kontext [dev]?

For a standard local Diffusers setup, 28 steps is a starting point based on the default in the FluxKontextPipeline call signature. As of October 2026, the linked Diffusers implementation also sets num_inference_steps=28, although the parameter prose on the documentation page says 50. Set the value explicitly to avoid relying on that conflicting description.

In an already configured local pipeline, the call looks like this:

result = pipe(
    image=input_image,
    prompt="Add a small silver star pin to her left blazer lapel.",
    num_inference_steps=28,
).images[0]

This is a documented default, not a measured optimum for every edit. When comparing counts, keep the input, prompt, seed, size, and other settings fixed. Check whether the requested change succeeds and whether details you wanted to preserve move or change.

BFL's Kontext [dev] model card links to Diffusers and ComfyUI implementations. For a ComfyUI workflow or an accelerated checkpoint, follow that workflow's sampling settings. A local step count does not reveal the count used by hosted [pro] or [max].

What does one hosted Kontext request produce?

The example below uses one hosted editing request with no sampling-count override. The input is a fictional adult portrait generated with FLUX.2 [pro]. Kontext [pro] received this exact instruction:

Add a small silver star pin to the left lapel of her blazer. Keep her face, hair, pose, clothing and background unchanged.

Input portrait	Kontext [pro] result

Input: FLUX.2 [pro], 880 × 1168 pixels. Result: FLUX.1 Kontext [pro], 880 × 1184 pixels, one editing request. The silver star is on her left lapel, on the right side of the image.

The pin appears where requested, and the brown bob, navy blazer, white shirt, and gray background remain. The crop and lapel shape change, and facial details are redrawn despite the preservation instruction. This pair shows one completed edit; it cannot establish a sampling count or compare the quality of different step settings.

What can you change in hosted Kontext when an edit misses?

Revise the instruction or use a clearer input. For a misplaced pin, specify “on her left blazer lapel, below the collar.” For another attempt at the same change, upload the original again. To make a second change to a result you want to keep, upload that result as the next input.

Compare the original and result before continuing, especially around the eyes, hairline, and garment seams. The Kontext quality guide shows how to inspect those details at full size.

As of October 2026, Kontext [pro] costs 4 credits per image on flux-3image, and [max] costs 7. Both accept one reference and offer the 1K output tier. If you need multiple references or 2K or 4K output, the FLUX 3 Image editor supports up to 10 references and those sizes. These are input and output options; they do not add a Kontext sampling-step control.