FLUX 3 Image is out: what to know about the release

FLUX 3 Image arrived on October 1

Black Forest Labs released FLUX 3 Image on October 1, through its API and the browser Playground. It's the still-image half of FLUX 3, the model BFL announced on July 23 as one system trained on images, video and audio together. The video side went generally available on August 4, so image users have been waiting eight weeks for this.

What it does

FLUX 3 Image handles text-to-image, edits to an existing image, and compositions built from up to 10 reference images.

The headline feature is layout control

Alongside the prompt, FLUX 3 Image accepts a JSON list of named elements. Each one gets a bounding box on a 0 to 1000 grid, written as [top, left, bottom, right], and a tag saying whether it's new, stays where it is, or moves. References are called by name (ref_image_0 through ref_image_9), so a single request can put the jacket from one reference on the person in a specific box.

For product shots, ad layouts and storyboards, that means placing the headline space or the product where you want it instead of rerolling until it lands there.

The same structure drives editing. BFL's docs show recolor, replace and move operations, and its launch examples, as reported by TechTimes, show how much of each image stayed pixel-identical after an edit:

How much of the image survives an edit

These are BFL's own showcase edits, so read them as a best case. BFL also says the boxes guide placement and scale rather than acting as hard limits, and elements can extend past them.

What it costs

A 50% launch discount runs until 8 a.m. Pacific on October 8. Prices per image, from BFL's pricing page and Kingy.ai's launch breakdown:

The 4k tier costs more than 12 times the 1k tier, so batch work at full resolution adds up fast: 1,000 images at 4k list price comes to $607.

What's still missing

There are no open weights yet. BFL says a public open-weight release is coming "within weeks" and hasn't given a date. It also hasn't published a parameter count or a GPU memory requirement, so anyone running FLUX locally in ComfyUI can't plan hardware yet. FLUX.2 [dev] remains the local option for now.

There are no independent rankings yet

On launch day, FLUX 3 Image had no entry on the Artificial Analysis text-to-image or editing leaderboards. Its predecessor had ground to make up. As of October 1:

FLUX.2 [max] trailed OpenAI by 176 points on text-to-image. BFL argues that training on video and audio gives FLUX 3 a better model of how physical scenes fit together, and it claimed top Elo scores for FLUX 3 Video in August. Nobody has shown yet whether that carries over to still images.

Should you try FLUX 3 Image now?

If your work is layout-heavy, the launch week is the cheap time to find out whether the bounding boxes hold up on your own briefs. A few dozen test images at 1k costs under a dollar until October 8.

If you run models locally, there's nothing to download yet. And if all you want is the best single image from a plain text prompt, it's worth waiting a couple of weeks for the leaderboards to place it.