Flux Kontext Prompts for Editing Existing Images: 8 Tested Examples
8 production-tested Flux Kontext prompts for editing existing images, with the formula, observed results, and what to avoid. From the 8frame canvas.
Flux Kontext prompts for editing work differently from generation prompts. The model reads your reference image first, then applies only the changes you describe. You don't rebuild the scene from scratch. You push on specific elements while the rest stays locked. Here are 8 prompts we ran on real images in the 8frame canvas, with the exact text and what each one produced.
TL;DR
- Flux Kontext is a reference-driven model: describe the change, not the whole image
- Specify what stays the same as explicitly as what changes
- Identity, lighting direction, and background geometry survive edits better than color and texture
- Each prompt below is verbatim from an 8frame session
When to use Flux Kontext for editing existing images
Kontext is the right pick when the base image is solid and you want surgical changes: swap a studio background for a city street, add a product to a lifestyle shot, shift noon light to golden hour. It does those without rebuilding from scratch.
It's not the right pick for compressed, low-resolution source images. The model struggles to separate subject from background at the pixel level. For those cases, see nano banana vs seedream vs flux to match the right model to the source quality first.
The prompt formula
The most reliable structure is: [Keep anchor] + [Target change] + [Technical constraint]
Keep anchor: what must not change. "Keep the subject's face, expression, and jacket unchanged" outperforms "keep the person the same."
Target change: the modification, precisely. "Replace the white wall with a rain-slicked Tokyo street at night, neon reflections in the wet pavement" beats "change the background to something urban."
Technical constraint: lighting match, aspect ratio, or output format. Kontext respects these when stated explicitly.
8 tested prompts for editing existing images
1. Background swap, subject preserved
Prompt: Keep the subject's face, hair, jacket, and pose exactly as in the reference image. Replace the plain grey studio background with a sun-lit alpine meadow, mid-afternoon light matching the existing front-lit exposure on the subject. No shadow inconsistency between subject and new background.
Subject stayed locked. The meadow came in with realistic depth of field and soft-out-of-focus grass at the edges. Shadow direction on the subject matched the new background lighting without manual adjustment. This is the most reliable Kontext category when you anchor the subject explicitly.
2. Lighting: time-of-day change
Prompt: Reference image: portrait of a woman standing on a city sidewalk, shot at noon with flat overhead light. Shift the lighting to golden hour: warm amber directional light from camera-left, soft fill from a reflector on the right, long shadow on the pavement behind her. Keep her clothing and expression unchanged.
Color temperature shifted on both subject and environment. Shadows landed where described and the building picked up warm ambient bounce. The only weak spot: hair went orange rather than warm gold. Adding "hair retains natural color, only lit by warm ambient" in a second pass corrected it.
3. Outfit change
Prompt: Keep the subject's face, skin tone, and pose exactly as in the reference. Replace the grey hoodie with a fitted dark navy blazer, white dress shirt visible at the collar, no tie. Maintain the original background. Fabric should have a natural matte texture, not shiny.
The blazer replaced cleanly and the matte fabric call-out worked. Without it, previous runs produced a silky look that read as costume. One watch-out: forearms can introduce edge artifacts where sleeve length changes. Adding "sleeves extend to the wrist" eliminated the floating-forearm artifact on the first run.
4. Object removal
Prompt: Remove the coffee cup from the table in the lower-right corner of the frame. Fill the space with the same wooden table surface, matching grain direction and ambient light. Do not alter anything else in the frame.
Removed cleanly in one pass, with wood grain infill that matched the surrounding surface. Because Kontext reads the whole image for context, the filled area inherits correct grain and shadow rather than producing a blurry patch. Reliable for objects under roughly 20% of the frame. Larger removals need a dedicated inpainting step.
5. Object addition
Prompt: Add a black matte espresso cup on the wooden table in the lower-left foreground, empty, no lid, steam rising gently. Cup should cast a soft shadow consistent with the diffuse overhead light already in the scene. No changes to anything else.
Cup appeared with correct shadow direction and subtle steam. The variance was size: it came in too large on the first two runs. Adding "cup diameter approximately 3 inches in frame-relative scale" stabilized it across three runs. Shadow direction matched the scene without extra prompting.
6. Color regrade
Prompt: Apply a desaturated, cool-toned grade to the entire image: lift the shadows toward cool blue-grey, pull the highlights toward pale white with a slight cyan tint, reduce overall saturation by about 30%. Preserve skin tone detail, slightly warmer than the surroundings. No change to composition or sharpness.
Grade landed close on the first try. Skin tones stayed warmer than the background as specified, avoiding the flattened-face look common in heavy desaturations. The "about 30%" language gave the model room to interpret rather than an exact percentage it can't control. For tighter color work, 8frame's image editing workflows include a Kontext color chain with side-by-side comparison before committing.
7. Style transfer onto existing photo
Prompt: Apply a hand-painted oil painting style to the reference photograph, as if rendered by a contemporary realist painter: visible brushstrokes, especially in the sky and background foliage, slightly thickened paint texture. Keep the subject's facial structure and likeness recognizable, face rendered with finer detail than the background. Maintain the original color palette, do not shift hue.
Brushstrokes distributed across the image with finer detail on the face as requested. Palette stayed close to the original. The phrase "facial structure and likeness recognizable" is load-bearing: without it, the face simplifies into a generic painted impression rather than the specific person in the reference.
8. Composition extension (outpainting)
Prompt: Extend the canvas to the left by approximately 40% of the current image width. Fill the new area with the same indoor environment visible in the original: matching wall color, light direction from camera-right, floor visible at the bottom of the frame. No subject should appear in the new area. The seam between old and new canvas should be invisible.
Extension came in with consistent wall color and light direction. Seam blended well on images with soft, low-frequency backgrounds (plain walls, blurred foliage). Visible on images with strong vertical lines or patterns at the frame edge. A 10-pixel feather mask before passing to Kontext removed the seam on a second run.
Common failures and fixes
Subject bleeds into background change. Add "do not alter the subject's color, texture, or outline" to the keep anchor.
Lighting inconsistency after background swap. State the existing key light direction explicitly and require the new background to match it.
Object additions appear too large or too small. Include a size reference relative to other objects in the frame rather than an absolute dimension.
Style transfer loses identity. Add "maintain facial likeness and recognizable features, apply style more heavily to background than subject."
Outpaint seam on patterned backgrounds. Feather the edge mask and add "the seam should be invisible, background pattern continues uninterrupted."
Step-by-step on 8frame
- Open the image editing workflow and upload your reference image to the canvas.
- Select Flux Kontext from the model picker.
- Write your prompt using the keep anchor + target change + technical constraint structure.
- Run the first pass. Use the canvas comparison view to evaluate before committing.
- For iterative edits (outfit change followed by background swap, for example), chain two Kontext nodes. Separate tasks produce cleaner separations than stacking both edits in one prompt.
- Export at native resolution or pass to an upscaler node for print-ready output.
FAQ
What makes Flux Kontext different from regular image generation prompts?
Kontext reads the reference first and builds the output as a modification. You describe the change relative to what exists. Subject identity and scene geometry carry over without being re-specified.
Can Flux Kontext change someone's face or identity?
Yes, if you don't anchor it. Phrases like "keep the subject's face, expression, and skin tone unchanged" lock identity during non-face edits.
How many edits can I stack in one prompt?
Two or three related changes usually work fine. Background swap plus lighting is fine. Background swap plus outfit plus style transfer plus outpainting in one pass produces inconsistent results. For complex sequences, chain separate Kontext nodes rather than stacking everything into one prompt.
For a head-to-head comparison of Flux Kontext against Nano Banana and Seedream on image generation tasks, see nano banana vs seedream vs flux.