Flux 3 Edit Prompts: 12 Edits That Keep the Rest Intact

Asking for an edit is easy. Getting back the same photo with exactly one thing different is the hard part. You ask for a green mug and the barista's face shifts, the window moves, the logo on the apron melts. Most bad edits aren't the model's fault. The prompt said what to change and never said what to keep.

These Flux 3 edit prompts fix that. Each one is written for Flux 3 Edit, the Black Forest Labs model that takes one to ten reference images plus an instruction and changes, restyles or combines them. There are twelve below, grouped by job: object swaps, backgrounds, relighting, text, style, product shots and multi-reference composites. Every prompt is copy-ready, and each comes with one line on why it holds the rest of the picture together.

For the full grammar (captions, the 0 to 1000 grid), see how to write better Flux 3 prompts. This is the cookbook next to it.

The Pattern Behind Every Good Flux 3 Edit Prompt

Good edit prompts have three parts, and most people only write the first. Say what changes. Say what stays. Say how the new thing should sit in the old light. That's it.

BFL's docs push hard on the first part: "Name the element precisely enough that only one thing in the image matches." If two things could match your noun, the model guesses, and that's how the wrong cup turns green.

The second part is where boxes come in. For a single change you don't need them. BFL says the model expands a plain instruction and may draw its own box. When you need certainty about what moves and what doesn't, you add a JSON list after the prompt: each row has an id, a from (which reference it comes from), a src_bbox (where it is now), a tgt_bbox (where it should be) and a desc. Boxes are [top, left, bottom, right] on a 0 to 1000 grid. Same source and target box means keep. A null source means generate something new. A null target means remove.

The third part separates a composite from a sticker. Shadows and nearby light can shift even outside your boxes, so say which way they go.

Object Swaps: Change One Thing by Name

1. Recolor one object with an exact color.

Change the color of only the ceramic mug on the left side of the table to #2F6B4F, matte glaze. Keep the second mug, the hands, the table and the window exactly as they are.

Why it works: "only" plus a position makes the target unique, and a hex code (BFL's docs use them too) is a color, not a mood.

2. Replace an object in a fixed spot.

In <ref_image_0>, replace the paperback book <book_1> on the bedside table with a small brass alarm clock, and keep the reading lamp <lamp_1> untouched. [{"id": "book_1", "from": null, "src_bbox": null, "tgt_bbox": [560, 120, 760, 330], "desc": "small brass twin-bell alarm clock with a cream dial, facing the camera"}, {"id": "lamp_1", "from": "ref_image_0", "src_bbox": [180, 90, 560, 360], "tgt_bbox": [180, 90, 560, 360], "desc": "linen-shaded reading lamp with a turned wooden base"}]

Why it works: BFL treats a replacement as a new row (no source, a target box), and the lamp gets a keep row because it's the thing most likely to drift when something beside it changes.

3. Remove an object and what it leaves behind.

In <ref_image_0>, remove the orange traffic cone <cone_1> and its shadow from the sidewalk, and fill the area with the same grey paving stones. [{"id": "cone_1", "from": "ref_image_0", "src_bbox": [610, 420, 880, 540], "tgt_bbox": null, "desc": "orange traffic cone with a white reflective band"}]

Why it works: a null target box is BFL's remove instruction, and the docs say to name what the object leaves behind, such as its shadow, so it goes too. Say the removal in the sentence as well as the row.

Background Swaps That Leave the Subject Alone

Unanchored background edits fall apart because the subject is the biggest thing in frame, and the model is tempted to re-render it too. Pin it.

4. New location, same person.

In <ref_image_0>, replace the office background with a sunlit rooftop terrace with potted olive trees and a soft city skyline, and keep the woman <person_1> exactly as she is, same pose, face, hair and clothes. Let the warm afternoon light fall on her from the right to match the new scene. [{"id": "person_1", "from": "ref_image_0", "src_bbox": [120, 300, 1000, 720], "tgt_bbox": [120, 300, 1000, 720], "desc": "woman in a rust linen blazer, shoulder-length dark hair, smiling at the camera"}]

Why it works: the keep row protects the subject's box, the sentence lists the features you care about, and the last line tells the light on her to change on purpose instead of by accident.

Relight Without Repainting: Flux 3 Image Edit Prompts for Light

BFL's docs don't list relighting as its own mode. It's an instruction like any other, and their multi-reference example asks the model to "adjust the lighting on the van to integrate well within image 1". That sentence is a template. Light is something you can ask for directly.

5. Move the time of day.

Relight Image 1 as golden hour: low warm sun from the left, long soft shadows across the courtyard, a faint orange rim on the fountain's edge. Keep every object, its position and the camera angle unchanged.

Why it works: a source, a direction and shadow behavior give the model a physical change, not a filter. The last sentence locks the geometry.

6. Make a composited object belong.

Place the leather backpack from Image 2 on the wooden bench in Image 1, and adjust the lighting on the backpack to match Image 1: cool overcast daylight from above, a soft contact shadow under the bag, no hard highlights.

Why it works: it follows BFL's own pattern (put the thing in, then fix the light on the thing), and the contact shadow is the detail that stops an object from floating.

Text Edits: Quote the New Words

BFL says Flux 3 renders high-accuracy text in multiple languages, and its bounding-box docs give one rule for text: each line gets its own row, with the exact words in quotes inside desc.

7. Change the words on a sign.

In <ref_image_0>, change the text on the café chalkboard <board_text> to read "TODAY: FIG TART", keep the chalkboard frame <board_1> and its hand-drawn border unchanged. [{"id": "board_text", "from": null, "src_bbox": null, "tgt_bbox": [300, 260, 460, 740], "desc": "white chalk hand lettering reading \"TODAY: FIG TART\", slightly uneven strokes"}, {"id": "board_1", "from": "ref_image_0", "src_bbox": [220, 200, 820, 800], "tgt_bbox": [220, 200, 820, 800], "desc": "black chalkboard in a weathered oak frame with a hand-drawn vine border"}]

Why it works: the new copy is a new row with quoted words, the board is anchored, and "slightly uneven strokes" keeps the same hand.

8. Translate a label.

In Image 1, replace the English words "SEA SALT" on the jar label with the German "MEERSALZ" in the same bold condensed sans serif, same cream color and same size. Keep the jar, the lid, the label's illustration and the background unchanged.

Why it works: you quote both the old and the new words so there's no doubt which text is meant, and you describe the type so the replacement inherits the design instead of the model's default font.

Style Changes That Keep the Composition

A restyle touches the whole image, so the layout can drift with the look. Name the structure that must survive.

9. Restyle a photo.

Restyle Image 1 as a hand-pulled linocut print in three inks (deep blue, brick red and off-white), visible carving marks and slight ink texture. Keep the composition, the cyclist's pose, the bridge and the horizon line exactly where they are.

Why it works: a named medium with its physical traits (inks, carving marks) beats "artistic", and the keep sentence names the four things a viewer would notice moving.

10. Borrow a style from a second image.

Repaint Image 1 in the style of Image 2: its loose watercolor washes, granulated paper texture and muted sage and ochre palette. Keep the subject, layout and framing of Image 1. Use Image 2 only for style, not content.

Why it works: BFL's advice for multi-reference is to say what each image contributes, and the last line stops objects from Image 2 sneaking into the scene. Image 1 goes first, so with aspect_ratio on auto the output keeps its shape.

Product Shots From One Packshot

One clean packshot, many scenes. The product has to come through untouched, so name the details that make it recognizable.

11. Put the product in a lifestyle scene.

Place the amber glass serum bottle from Image 1 on the marble bathroom counter in Image 2, standing upright to the right of the sink. Keep the bottle's shape, dropper cap and label exactly as in Image 1. Match the soft window light from the left in Image 2, with a faint reflection of the bottle in the marble.

Why it works: the identifying parts are listed, the placement is specific, and the reflection makes the bottle sit on the stone, not above it. Since the scene is Image 2, set aspect_ratio yourself. For an invented brand, try grounding off: it runs web and image search before generating, and a made-up name has nothing to find.

Multi-Reference Edits: Give Every Image a Job

Flux 3 Edit takes up to ten images in image_urls, each 256 by 256 pixels to 16 megapixels. More references don't make a better composite. Clear jobs do.

12. Furnish a real room from separate product shots.

Furnish the exact room in Image 1: the green velvet sofa from Image 2 against the back wall under the window, the round jute rug from Image 3 on the floor in front of it, and the brass floor lamp from Image 4 to the right of the sofa. Keep the walls, window, floorboards and camera angle of Image 1 unchanged. Match all three items to the room's warm morning light.

Why it works: it's built on BFL's own "furnish the exact room" example, every image has one role, each item has a place relative to something already in the room, and one closing line handles the light for all of them. If the sofa lands in the wrong spot, add a row with "from": "ref_image_1" and a target box rather than another sentence.

Chain Flux 3 Edit Prompts, One Change per Pass

Six changes at once is the fastest way to wreck a good image, because you can't tell which one broke the face. Edit in passes: recolor, check, swap the background, check. Each output becomes Image 1 of the next request, and your boxes keep working as long as the aspect ratio holds.

Settings matter less than prompts, but three are worth knowing. Run drafts at 1k (the default) and only go to 2k or 4k for the keeper; the schema warns that 4k can take several minutes. Leave aspect_ratio on auto unless the base image isn't first. And give new elements room: BFL found that tiny boxes, around 40 by 25 pixels, often didn't produce the element at all. Boxes guide, they don't clip, so a little slack is fine.

No base image yet? Make one in Flux 3 Text to Image and hand it to Edit. Both sit in the Flux 3 family on each::labs, and each::labs flows chain them with other steps. Painted masks are still a job for Flux Fill Pro; Flux 3 wants words and boxes. BFL's full reference is in its FLUX 3 layout docs, more image-to-image models are a click away, and our Flux 3 Image explainer covers what changed since Flux 2.

Frequently Asked Questions

Do Flux 3 edit prompts need bounding boxes?

Only when certainty matters. For one clear change a precise sentence is enough, since the model may add its own box. Add rows when several things change, when something nearby must not move, or when the first try hit the wrong object.

How do I refer to my images in a Flux 3 Edit prompt?

Upload order sets the numbers. In prose, call them Image 1, Image 2 and so on; in the JSON rows, BFL's docs write the first reference as ref_image_0, so rows count from zero. Put the image whose framing you want first, since auto aspect ratio follows it.

Why did the lighting change outside my box?

Because light doesn't respect rectangles. Pixels outside the boxes typically stay unchanged, but BFL notes that shadows, reflections and nearby lighting can still shift when you add or move something. Describe the light you want in the prompt, and anchor the objects you care about with keep rows.

Can Flux 3 Edit change text in an image?

Quoting is the whole trick. Put the new words in quotes, give each line its own row if you use boxes, and describe the lettering so it matches. Short copy in a roomy box lands best.