FLUX 3 Image Edit

FLUX 3 Image Edit — AI Image Editing and Composition (Image-to-Image)

What is FLUX 3 Image Edit?

FLUX 3 Image Edit is Black Forest Labs' frontier image editor and the image pillar of the multimodal FLUX 3 family. Hand it one image and a plain-language instruction, or up to ten reference images, and it edits, restyles or combines them. Its defining trait is pixel-exact control: recolor, replace, move or remove a single element and every pixel outside that region stays unchanged, so repeated multi-turn edits never degrade the rest of the frame.

Key Features

• •Pixel-precise local edits — change one element and keep everything else identical to the source.
• •JSON bounding boxes written as [y_min, x_min, y_max, x_max] on a 0-1000 grid to place, move, recolor or replace a region; a language model can author the layout from a short brief.
• •Up to ten reference images combined in one request — objects, products, clothing, faces or styles.
• •Whole-image restyling that keeps the original composition.
• •Native text-and-web grounding for real places, products and people.
• •Native 4K (16 MP) output in a single pass with no separate upscaler; tiers 768sq, 1k, 1.5k, 2k, 4k and aspect ratios from 21:9 to 9:21.

Best Use Cases

In testing, a cherry-red motorcycle fuel tank was recolored to brushed copper while the pegboard of tools, chrome and brick wall stayed pixel-for-pixel identical — exactly the surgical edit that e-commerce, advertising and product-design teams need. Swap a product color, replace a single object, move an element to new coordinates, composite a product into a scene from a second reference, or restyle a whole frame into a blueprint or illustration. Because boxes are normalized, one layout description works at any output size, which makes the model a strong fit for multi-step pipelines and LLM agents.

Prompt Tips and Output Quality

Name the element and state what must stay: 'change X ... keep everything else exactly unchanged'. For placement, tag the element with an <id> in the prompt and append the bbox array. Refer to inputs as Image 1 and Image 2 and say which supplies the object and which is the scene. In testing, 1k edits returned clean results in roughly 50-90 seconds; step up to 2k or 4k when you need print detail.

How to Use FLUX 3 Image Edit

FLUX 3 Image Edit turns a plain-language instruction into a precise change on an image. The only required inputs are one to ten reference images and a prompt. Phrase the prompt as what to change plus what must stay — add 'keep everything else exactly unchanged' to lock the rest of the frame — and the model edits only the region you name.

Recommended Settings

• •resolution — 1k (about 1 MP) for web and fast iteration; step to 2k or 4k for print or large screens. Price scales with the tier.
• •aspect_ratio — leave on auto to keep the first reference image's framing, or pick a fixed ratio from 21:9 to 9:21.
• •grounding — on for real places, products or logos; off for invented scenes and faster results.
• •safety_tolerance — default 2 (0 is strictest, 4 most permissive).
• •bounding boxes — tag an element with an <id> in the prompt and append a JSON array of {id, bbox:[y_min, x_min, y_max, x_max] on a 0-1000 grid, desc}.

Example edits to try

Each edit below ran end to end on a single garage photo at the settings above; everything outside the instruction carried over unchanged.

• •Recolor, everything else locked. Prompt: "Recolour only the motorcycle's fuel tank from cherry red to brushed copper. Keep the chrome, the tyres, the pegboard of hand tools, the brick wall and the lighting exactly unchanged." Only the tank turned copper while the busy pegboard and chrome stayed identical. Tip: name the element and add 'keep everything else exactly unchanged' to protect the rest.
• •Move by bounding box. Prompt: "Move the red helmet from the motorcycle seat onto the empty left side of the workbench top; leave the seat empty. Keep the motorcycle, the pegboard of tools, the wall clock and the brick wall exactly unchanged. <helmet_1> [...bbox...]" The helmet relocated to the exact bench coordinates with no duplicate. Tip: write boxes as [y_min, x_min, y_max, x_max] on a 0-1000 grid so one layout works at any resolution.
• •Restyle, composition kept. Prompt: "Restyle the whole image as a hand-drawn cutaway blueprint: thin white technical lines on deep blue drafting paper. Keep the motorcycle, the workbench, the pegboard tools and the helmet in their original positions and proportions." The whole frame became a white-line blueprint with every part in place, down to a cutaway of the engine. Tip: name the target style and add 'keep every part in its original position' to restyle without re-composing.

Reference-guided composition

FLUX 3 Image Edit takes up to ten reference images; address them as Image 1, Image 2 and say which supplies the object and which is the scene.

• •Combine two references. Prompt: "Hang the framed poster shown in Image 2 on the empty brick wall to the left of the pegboard in Image 1. Keep the motorcycle, the pegboard of tools, the helmet and everything else in Image 1 exactly unchanged." The poster was composited onto the wall with matching perspective and shadow while Image 1 stayed untouched. Tip: the first image is the scene, the second supplies the object to place.

Built for more besides: try "Replace the seat with a quilted black one, keep the rest unchanged", "Add a white painted shop sign reading GARAGE on the brick wall", or "Change the floor to checkered tiles and keep the motorcycle exactly as it is".

How are bounding boxes written?

As [y_min, x_min, y_max, x_max] on a normalized 0-1000 grid, so one layout works at any resolution.

How many reference images can I use?

Up to ten per request, addressed as Image 1, Image 2 and so on.

Does it really keep the other pixels unchanged?

Yes — regions outside your instruction stay unchanged, which is what makes repeated edits safe.

Is 4K output native?

Yes, FLUX 3 Image Edit renders up to 4K (about 16 MP) in a single pass without a separate upscaler.

What does grounding do?

It lets the model check web and image search for real references before rendering; turn it off for pure imagination.