Skip to content
AI Primer
release

FLUX 3 Image adds bounding-box layout controls

FLUX 3 Image adds bounding-box layouts, precise multi-turn edits, 4K output and support for up to 10 references. A creator comparison found it most accurate at following fine-detail editing instructions.

4 min read
FLUX 3 Image adds bounding-box layout controls
FLUX 3 Image adds bounding-box layout controls

TL;DR

A camel becomes a futuristic vehicle inside BFL's editing demo, while the desert backdrop stays put. One creator combined a kneeling pose and peppermint-camo trousers in a single edit prompt; the 4K tier in Runware's docs reaches 4096 × 4096 for square images.

Bounding-box layouts

BFL supports bounding boxes both when composing an image from scratch and when editing an existing one.

Runware's API reference exposes the controls through settings.boundingBoxes:

  • id and description: a unique region identifier and a description of its contents.
  • referenceIndex: the input image supplying an element, indexed from 0 to 9. A null value denotes a new element.
  • sourceBox and targetBox: coordinates written as [top, left, bottom, right] on a normalized 0-to-1000 grid.
  • Matching coordinates: keep the element in place. Different source and target coordinates move or resize it.
  • targetBox: null: removes an element from a supplied reference image.

The coordinate scale stays the same regardless of output dimensions.

Pixel-preserving edits

BFL promises precise edits “without changing any other pixel.” Its replacement demo targets a camel inside a desert scene.

The API documentation uses softer wording: pixels outside an editing frame “generally” stay unchanged, according to Cocoloop's launch report.

Kimono recoloring and macro close-ups

Creator ai_artworkgen compared editing instructions across Midjourney, Ideogram 4.5 and FLUX 3 Image.

The kimono test used three requests:

  1. “Change her kimono to pink silk with white flowers, and a custard yellow bow.”
  2. “Show me a macro closeup of the subject's face.”
  3. “Show me a macro closeup of the subject's hands on her kimono.”

Results differed on both instruction following and continuity:

  • Midjourney: handled the two close-ups, but missed the white flowers and exact custard-yellow shade, according to ai_artworkgen.
  • Ideogram 4.5: matched the garment changes, but the macro shots' lighting looked “a bit off” to ai_artworkgen.
  • FLUX 3 Image: matched the garment changes and earned praise for lighting and contextual awareness in the close-ups, according to ai_artworkgen.

Chained edits and pose changes

Ideogram 4.5 also held up across successive edits, sometimes with two changes per iteration while maintaining image quality, ai_artworkgen reported in a follow-up.

For a separate FLUX test, the creator used a Midjourney original and reported that the model “nailed so many edits” on the first attempt. The prompt bundled a new pose with several wardrobe changes:

Marble-column edits

The final comparison requested: “Change some of the detailing on the marble columns to camo print.” FLUX was the only model that changed just the columns, according to ai_artworkgen.

The smallest edit is the best demo in this thread.

Ten reference images

A composition can draw from up to 10 reference images, according to BFL's launch thread.

Runware's reference-image schema accepts:

  • Image UUIDs.
  • Image URLs.
  • Data URIs.
  • Base64-encoded images.

Each reference must be at least 128 pixels in size. A bounding box's referenceIndex connects a placed element to the relevant input image.

4K output

Runware's dimension table puts the 4K tier at approximately 16.8 megapixels, with dimensions varying by aspect ratio:

  • Square: 4096 × 4096.
  • Landscape, approximately 16:9: 5456 × 3072.
  • Portrait, approximately 9:16: 3072 × 5456.

API prices and commercial weights

FLUX 3 Image is available through BFL's Playground and its API. BFL's API launch promotion offers 50% off until October 8, according to its launch thread.

Runware's published rates list discounted and standard per-image prices:

  • 0.75K: $0.0205 discounted, $0.041 standard.
  • 1K: $0.024 discounted, $0.048 standard.
  • 2K: $0.05 discounted, $0.10 standard.
  • 4K: $0.303 discounted, $0.607 standard.

Commercial weights are available under a license for fine-tuning and deployment on a company's own infrastructure. BFL says an open-weight version will launch “in the coming weeks” in the same announcement.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 4 threads
TL;DR1 post
Kimono recoloring and macro close-ups2 posts
Chained edits and pose changes1 post
4K output1 post
Share on X