● Gemini Omni Prompts
Home / All Prompts / object-swap

Object Swap & Background Replace via Conversational Editing

Iteratively swap objects and replace backgrounds using Gemini Omni's flagship conversational editing — keep characters and lighting stable across turns.

object-swap 16:9 10s Conversational edit #conversational-editing#omni-flagship#swap#background#iterative

Prompt

Turn 1 — Generate base scene:
A young woman in a white linen dress stands in a sunlit Victorian parlor, holding a porcelain teacup. Camera at chest height, soft window light from the left, locked-off shot, 10 seconds.

Turn 2 — Swap one object:
Change the porcelain teacup to a small bouquet of wildflowers. Keep her posture, the parlor, and the lighting exactly the same.

Turn 3 — Replace the background:
Now replace the parlor background with a misty forest clearing at dawn. Keep her, her dress, and the bouquet identical. Match the original key light direction (upper-left).

Why this prompt

Conversational editing is Gemini Omni’s flagship differentiator vs Sora / Veo. Google’s DeepMind prompt guide explicitly shows the model handling iterative object swaps (butterfly → bee → fireflies, astronaut → sea anemone) where each turn builds on the previous result.

This 3-turn structure is adapted from Atlas Cloud’s hands-on review, which validated the exact pattern: generate base → swap one variable → replace another.

Source tier: 🟢 Official demo + media hands-on (high confidence)

The magic incantation: “Keep [X] exactly the same”

Seaart’s prompt collection and Google’s own documentation both emphasize: vague editing prompts cause over-edit drift. When you swap one thing, explicitly list what you want preserved.

“Replace the food on the plates with creamy pumpkin soup. Keep the two people talking, their movements, facial expressions, ocean background, and all lighting exactly the same.” — Seaart prompt collection

How to tweak

  • Single-variable rule: change only ONE thing per turn. Multi-edit prompts confuse Omni.
  • Lock list: explicitly preserve face / clothing / posture / lighting / background
  • Iterative chain: each turn references the result of the previous turn (not the original)
  • Reset point: if drift accumulates after 4-5 turns, regenerate from a fresh base

Common failure modes

Per TechCrunch’s launch report and Atlas Cloud’s testing:

  • Without “Keep X identical”, Omni may re-style the subject’s clothing or face when changing the background
  • Multi-object swaps in one turn drift unpredictably
  • Fine facial details and hand articulation degrade across 4+ turns (atlascloud scored character consistency 3/5)

Notes

  • Requires Gemini Omni Flash subscription (Gemini app or Google Flow) for multi-turn editing
  • All output includes Google SynthID invisible watermark — by design, not removable
  • Single clip max 10 seconds (Flash hard limit, per TechCrunch + CineD)

Sources