How to edit photos with FLUX.2: 10 practical recipes

Runbo Li
Runbo Li
·
· 6 min read
Edit Photos With Flux Kontext

Quick answer

To edit a photo with FLUX.2, upload the source image, describe the finished state you want, and identify what must remain recognizable. Start with one change, compare the output with the original, then make the next edit. For new projects, Black Forest Labs recommends FLUX.2; FLUX.1 Kontext is now a previous-generation path.

A useful first prompt is: “Change the wall to pale blue. Keep the person, furniture, framing, lighting direction and camera perspective unchanged.” This guide gives ten patterns you can adapt. They are prompt recipes, not results from a controlled Magic Hour benchmark.

Edit one detail in your image

Upload an image, describe one specific change, and compare the result with the original before making another edit.

Try AI Image Editor

FLUX.2 or FLUX.1 Kontext?

  • Choose FLUX.2 for a new workflow. The API supports single- and multi-reference editing, up to eight references depending on total pixels, and output up to 4 megapixels. The BFL playground can accept up to ten references.
  • Keep FLUX.1 Kontext only when an existing provider or integration is pinned to it. BFL describes Kontext as previous-generation; its direct API accepts one input image up to 20 MB or 20 megapixels and normally returns about a 1-megapixel output.
  • Check the interface you actually use. A third-party platform can expose different model versions, input counts, prices, aspect ratios or controls from BFL's own API and playground.

BFL currently offers FLUX.2 variants for different jobs: [pro] for production scale, [max] for maximum precision, [flex] for fine control and [klein] for cost-sensitive high-volume use. Availability and limits differ by endpoint, so record the exact model rather than writing only “FLUX.”

A reliable editing workflow

  • 1. Preserve the original. Work from a copy and record its dimensions, color profile and intended final use.
  • 2. Name each reference. With multiple inputs, state which image supplies the subject, product, pose, style or background.
  • 3. Describe the desired result positively. BFL's FLUX.2 [pro]/[max] guide says negative prompts are not supported. Ask for “sharp lettering” instead of “no distorted text.”
  • 4. Change one important thing first. A narrow first pass makes it easier to identify why an edit failed. Add secondary styling only after the subject and composition hold.
  • 5. Compare against the source. Inspect identity, product shape, labels, hands, edges, perspective, light direction and any element the prompt said to preserve.
  • 6. Save the model and prompt. For API work, also retrieve the result promptly: BFL says its signed result URLs are valid for ten minutes.

A prompt template that is easy to debug

Change: [the exact object, text, background or lighting]. Use: [reference image and its role, if any]. Keep: [identity, product geometry, pose, camera, lighting or layout]. Finish as: [realistic photo, product image, poster, editorial illustration or other deliverable].

Do not add every clause by default. Use only the constraints that matter to the asset. If the output is close, change one instruction rather than rewriting the entire prompt.

10 practical FLUX photo-editing recipes

1. Change a background and preserve the subject

Replace the background with a bright modern office. Keep the person, pose, clothing, hair, camera angle and light direction unchanged. Match the new room's perspective and shadows to the subject.

Workflow recipe output preview: Liquid Glass Product Grid

Workflow recipe

Subject transforms into a premium 16:9 liquid glass Bento infographic for [insert product name] in [English]. Hero color derived from product’s natural dominant color. Full saturation for product and accents, muted (30–40%) for icons and borders. Apple-style liquid glass cards (85–90% transparent), whisper-thin borders, subtle shadows, floating depth. Background [ethereal / macro / pattern / context], heavily blurred behind cards, with soft motion effect. Asymmetric Bento layout: hero ~30%, others ~70%. M1 — Hero: Subject displayed as [real photo / 3D glass / stylized], elegant composition + product name M2 — Core Benefits: 4 benefits + icons M3 — How to Use: 4 methods + icons M4 — Key Metrics: 5 exact data points (FOOD / MEDICINE / TECH format) M5 — Who It’s For: 4 suitable + 3 caution groups M6 — Important Notes: 4 precautions M7 — Quick Reference: category-based summary M8 — Did You Know: 3 facts Ultra-clean, high-end, futuristic aesthetic, sharp typography, balanced spacing, visually premium.

Model
nano-banana-2
Format
auto

Check: hair edges, contact shadows, reflected color and whether the face changed. For a simple transparent cutout, a dedicated background remover may be faster than generative replacement.

2. Place a product in a new scene

Prompt: “Place the bottle from image 1 on the stone counter in image 2. Preserve the bottle's exact silhouette, cap, label placement, logo, material and color. Match the counter's camera angle, reflections and daylight.”

Check: every letter, edge, cap dimension and reflection. Use the real product photo as a fixed layer when exact packaging cannot change.

3. Replace text on a sign or poster

Prompt: “Replace ‘OPEN TODAY’ with ‘OPEN FRIDAY’. Keep the same position, capitalization, type style, spacing, sign material, perspective and lighting.”

Check: spelling at full resolution. BFL recommends quotation marks around the original and replacement text for Kontext text edits; FLUX.2 adds improved text editing, but generated lettering still needs inspection.

4. Relight a portrait

Prompt: “Change the portrait to soft window light from camera left. Preserve facial identity, expression, skin texture, hair, clothing, framing and background. Use natural shadow falloff and neutral skin color.”

Check: eye color, face shape, earrings, hairline and whether the new shadows agree with the stated light direction.

5. Change clothing while preserving the person

Prompt: “Replace the jacket with the navy jacket from image 2. Keep the person from image 1, including face, body proportions, pose and hands. Fit the jacket naturally and preserve the original scene lighting.”

Check: fingers, garment openings, logos, overlaps and whether the edit changed body shape.

6. Remove one object and reconstruct the scene

Prompt: “Remove the red bag beside the chair. Reconstruct the floor and wall behind it using the surrounding texture, perspective, grain and lighting. Keep every other object unchanged.”

Check: repeated textures, bent floor lines, leftover shadows and changes outside the edited area.

7. Match a pose from another reference

Prompt: “Use the person and clothing from image 1. Match the standing pose and gaze direction from image 2. Keep the face, hairstyle and outfit recognizable. Preserve natural anatomy and place the person in the room from image 3.”

Check: hands, joints, face identity and whether each reference played the role you assigned. This is a FLUX.2 multi-reference workflow, not a single-input Kontext recipe.

8. Apply a visual style without losing the layout

Prompt: “Render image 1 as a hand-painted editorial gouache illustration. Preserve the subject positions, crop, silhouette and negative space. Use the muted red, blue and cream palette from image 2.”

Check: composition first. A strong style match is not useful if the product, person or layout moved.

9. Create a clean ecommerce image

Prompt: “Create a square ecommerce image of the product from image 1 on a warm white seamless background. Preserve the exact product shape, label and color. Center it with even margins, a soft contact shadow and diffuse studio lighting.”

Check: label fidelity, centering, scale across the product set and the marketplace's required output size.

10. Make a controlled ad variant

Prompt: “Keep the product, headline, logo, legal copy and layout from image 1. Change only the background from pale blue to hex #F4D35E and change the call-to-action button to hex #6C3BFF. Preserve readable text and WCAG-conscious contrast.”

Check: exact copy, brand colors, contrast and whether any legal text moved. Use a layout tool for the final typesetting when every character must be exact.

Diagnose the failure before rerunning

  • The subject changed: shorten the edit and explicitly preserve identity, pose, crop and distinctive features.
  • The references were mixed up: assign a numbered role to every input image in the prompt.
  • The edit spread too far: name the exact object or region and state which surrounding elements stay fixed.
  • Text is wrong: quote the exact old and new text, then inspect every character. Finish critical typography in a conventional editor.
  • The output is too small: choose a FLUX.2 endpoint and supported dimensions for the required deliverable, or use an image upscaler after checking that important details survive.
  • A later edit degraded an earlier one: return to the last accepted output and branch from it instead of accumulating damage.

Where Magic Hour fits

Use Magic Hour's AI Image Editor when you want a browser workflow for prompt-based changes without managing BFL API requests directly. Record the selected model shown in the interface, because platform availability can change independently from BFL's direct endpoints. For production decisions, compare the actual downloaded file rather than assuming every provider exposes the same FLUX variant or defaults.

Frequently asked questions

Yes. BFL still documents FLUX.1 Kontext [pro] and [dev], but calls Kontext previous-generation and recommends FLUX.2 for new image-generation and editing projects.

BFL's direct FLUX.2 API supports up to eight references depending on the combined input and output pixels; its playground supports up to ten. Variant and provider limits differ, so check the endpoint you will actually call.

BFL's [pro] and [max] prompting guide says no. Describe the desired visible state positively, such as “sharp subject with a clean empty background.”

No generative edit should be assumed perfect. Compare identity, geometry, logos, labels and small details with the source. Use the original asset or a conventional editor wherever exact preservation is required.

Official sources checked

Model status and limits were checked September 13, 2026 using BFL's FLUX.2 image-editing documentation, FLUX.1 Kontext image-editing documentation and FLUX.2 prompting guide. Limits, endpoints and recommendations can change; recheck the selected provider before a production batch.

Runbo Li
Runbo Li
CEO of Magic Hour
Runbo Li is the Co-founder and CEO of Magic Hour, where he builds AI video and image tools for content creation. He is a Y Combinator W24 founder and former Data Scientist at Meta, where he worked on 0-1 consumer social products in New Product Experimentation. He writes about AI video generation, AI image creation, creative workflows, and creator tools.
View author →

Continue Reading

Editorial comparison of precise image editing and style exploration with a legacy-to-current visual timeline
Recommended next
FLUX Kontext vs Midjourney V7: legacy and current choices

FLUX.1 Kontext and Midjourney V7 are legacy choices. Compare FLUX.2 and Midjourney V8.2 for editing, references, APIs, pricing and privacy.

Flux pro vs dev vs schnell
FLUX.1 Pro vs Dev vs Schnell: access, license and use
Side-by-side comparison of Flux.2 and Nano Banana Pro AI-generated images highlighting cinematic style versus logical accuracy
FLUX.2 vs Nano Banana Pro: editing, text, cost & access
Flux <> MH
FLUX.1 Kontext tutorial: edit images, prompts & FLUX.2
The Best AI Image Editors
6 best AI image editors in 2026: choose the right one