Add or Remove Objects

A photo → a clean edit with an object seamlessly added or removed.

Category: edit · Difficulty: Beginner · ~Time: 3 min

What you'll make

Targeted, instruction-driven edits to a single photo: delete a photobomber, erase a logo or wire, or add a prop that wasn't there. The model fills the gap with plausible texture and lighting so the result looks like the original shot, not a patch job.

Models used

🎨 Image — Flux Kontext (edit) (resolve live; e.g. black-forest-labs/flux-kontext-dev)
🎨 Image (alt) — Nano Banana 2 (edit) (resolve live; e.g. google/nano-banana-2/edit)
💬 LLM — Claude (to phrase precise edit instructions), optional

Steps

Upload the source photo — "Upload this image and give me the URL." → maps to atlas_upload_media; keep the download_url.
Write a precise instruction — describe exactly what to add or remove and what must stay untouched. Flux Kontext follows localized edit instructions well.
Run the edit — "Remove the [object] and reconstruct the background behind it." → maps to atlas_generate_image with Flux Kontext (edit), passing the uploaded URL as the input image.

Tip: ask the agent to atlas_search_docs for Flux Kontext first to confirm the live input-image field and any guidance/strength params.
Refine in passes — chain edits one change at a time (remove first, then add) for cleaner, more controllable results than one mega-prompt.

Prompt starters

Remove the person standing in the background on the right. Reconstruct the
brick wall and sidewalk naturally behind them. Keep everything else identical.

Add a steaming cup of coffee on the wooden table next to the laptop, with
soft morning light and a realistic shadow. Do not change the laptop or hands.

Tips & variations

Flux Kontext excels at "keep everything else the same" localized edits; switch to Nano Banana 2 (edit) when an add needs richer scene reasoning or relighting.
Removing large objects? Expect the model to invent what's behind — add "reconstruct plausibly" and re-roll if the fill looks off.
One change per run. Stacking "remove X and add Y and recolor Z" in a single prompt lowers fidelity.