Inpainting: Fixing Brand Images Without Reshoots

A skincare brand I work with rebranded their packaging in March. Their asset library — two years of shoots, roughly 400 hero images — showed the old label on every single one. The traditional quote for reshooting the top 60 images was five figures and six weeks. What actually happened: the new label was inpainted onto the existing photography over two working days, and the old shoots kept earning. Same lighting, same models, same compositions the audience already responded to. New packaging.

That's inpainting's job description: regenerate a selected region of an image while leaving everything else untouched. Mask the label, describe the replacement, keep the photograph. It's the least glamorous tool in the AI imaging stack and, for established brands sitting on photography libraries, probably the highest-ROI one. Here's the working craft — masking, prompting, and the edits you should refuse to make.

What inpainting is for (a taxonomy of fixes)

The brand-photography use cases, roughly ordered by how often I actually do them:

• Object removal. The extension cord in the lifestyle shot, the competitor's product in the cafe background, the exit sign, the passerby. Mask it, prompt the background continuation, gone.
• Packaging and label updates. The rebrand scenario above. Highest dollar value per edit.
• Prop swaps. The out-of-season flowers become evergreen; the dated phone in the model's hand becomes a current one; the breakfast in the hotel shot changes cuisine for a regional campaign.
• Blemish and defect repair. Scratched product surface, dented box corner, wrinkled backdrop, fingerprint on glass.
• Compliance edits. Removing a visible brand you don't have clearance for, a license plate, a recognizable face in the background.
• Seasonal redressing. Same hero composition, summer props swapped for holiday props — one shoot becomes a year of campaign variants.

The unifying economics: photography's expensive part is the shoot — lighting, talent, art direction. Inpainting lets you amortize that cost across changes the shoot never anticipated.

Masking is 80% of the skill

Everyone focuses on prompts; results are decided by masks. The rules I teach:

1. Mask generously past the object's edge. A tight mask forces the model to blend in a sliver of space, and halos live in slivers. Extend the mask 10-20 pixels beyond the object so the model rebuilds a natural transition zone.
2. Include the object's shadow and reflection in the mask. Removing a bottle but not its shadow leaves a ghost silhouette that screams "edited." Glass tables and glossy surfaces double this rule — the reflection is half the object.
3. Don't mask what must survive. If the region touches your product, respect the boundary pixel-precisely on the product side while staying generous on the background side. Asymmetric masks are normal.
4. Big edits in stages. Replacing a whole table setting works better as two or three sequential masked edits than one giant mask, which drifts toward the model redecorating the scene.

If you catch yourself masking more than about a third of the image, stop — you're no longer editing a photo, you're generating a new one, and the image-to-image workflow is the honest tool for that job.

Prompting the patch

An inpainting prompt describes what the masked region should become, in the context the model can see around it. Patterns that work:

Removal: "empty marble countertop continuing naturally, soft window light from the left, subtle reflections"

Swap: "a ceramic vase with white tulips, matching the warm afternoon lighting of the scene, soft shadow to the right"

Repair: "smooth undamaged matte black surface, continuing the product's curve, consistent studio lighting"

Three details carry the weight:

• Restate the light. The model sees the surrounding pixels, but naming the light direction and quality ("soft window light from the left") measurably improves shadow coherence on inserted objects.
• Describe the surface it sits on. Objects need contact — "soft shadow beneath, resting on the wooden table" prevents the floating-prop look.
• For removals, describe the background, not the removal. "No cord" invites cords. "Clean white backdrop, seamless" gets you the empty space. This matters enough that the prompt engineering guide treats it as a general law: models render nouns you mention, including negated ones.

Model choice note: precision-editing work rewards strong instruction-following models — Flux-family and Nano Banana 2-class editors handle "change only this, preserve everything else" markedly better than older checkpoints, and Versely routes edit tasks accordingly.

The label problem, honestly

Packaging swaps deserve their own section because they combine inpainting's best use case with its known weakness: text. Generative models still mangle small type — your ingredients list will come back as convincing gibberish.

The professional pipeline for a label swap:

Step	Tool	Why
1. Remove old label region	Inpaint	Rebuilds the container surface with correct curvature and light
2. Generate/prepare new label flat	Design file or text-to-image	Real typography stays vector-true
3. Composite label onto container	Editor (warp + blend)	Wrap to the surface curve, match highlights
4. Final harmonizing pass	Light inpaint over edges	Melts the composite's seams into the photo

Yes, it's four steps instead of one prompt. It's also the difference between a rebrand you can ship to retail partners and one that fails the first close look. For hero images where the label is large and legible, do it properly; for small-in-frame background products, a direct inpaint with the brand's colors is usually fine.

Edits to refuse

Where I draw lines, both practical and ethical:

• Don't inpaint claims into evidence. Adding a certification seal, an award badge, or a "before/after" difference the product didn't produce isn't editing — it's fabrication, and in regulated categories it's legal exposure.
• Be careful editing people. Cleaning a backdrop behind a model is routine; altering the model's body or face crosses into territory that several markets now regulate for advertising and that audiences increasingly punish.
• Don't fight fundamental light mismatches. If the replacement object needs opposite lighting to the scene, the composite will always feel wrong. Reshoot or regenerate the whole scene instead.
• Know when the photo is beyond editing. Three or more major edits in one image usually means the brief changed, not the image. Generate fresh — and if the background is the whole problem rather than any object in it, background removal plus a new backdrop is faster than inpainting the world piece by piece.

FAQ

What's the difference between inpainting and outpainting?

Inpainting regenerates a masked region inside the image — removing objects, swapping props, repairing defects — while everything outside the mask stays untouched. Outpainting extends the canvas beyond the original borders. They're sibling tools: one fixes within the frame, the other grows the frame.

Can inpainting really replace product reshoots?

For a large class of changes, yes: packaging updates, prop swaps, background cleanups, seasonal redressing, and defect repair on existing photography. It can't replace shoots that need new compositions, new talent, or fundamentally different lighting — the shoot's DNA is what you're preserving.

Why does inpainted text or labeling look wrong?

Generative models approximate small type, producing plausible gibberish. Never generate final label text directly. Inpaint the surface clean, composite real vector typography in an editor, then run a light harmonizing pass over the edges.

How do I avoid halos and ghost shadows in edits?

Mask generously — 10-20 pixels beyond the object's edge — and always include the object's shadow and any reflections in the mask. Tight masks and forgotten shadows are the two causes of nearly every "obviously edited" result.

Is it ethical to inpaint brand photography?

Removing distractions, updating packaging, and redressing scenes is standard practice — the honest core of the image is unchanged. The line is fabricating substantive claims: certifications, results, product capabilities. If the edit would change a buyer's factual understanding, don't make it.

Sitting on a photo library that almost works? Fix it instead of reshooting it — inpaint, edit, and generate in text-to-image. Free credits daily.