Why every AI editor re-imagines your photo — and how to stop it
The most common frustration in AI photo editing this year isn't quality — it's drift. You ask for a different background and the model returns a different person standing in it. Understanding why that happens makes it fixable.
What the model is actually doing
Generative editors aren't Photoshop with a text box — they rebuild the image. Given an instruction and your photo, the model re-renders the frame with the instruction applied, which means every pixel is technically regenerated. How much *moves* during that re-render is the whole question.
The difference between editors is what they're allowed to touch. Instruction-following models edit the described region and carry the rest over; weaker integrations re-imagine the whole frame. Uncontrolled regeneration is the failure mode the entire industry is currently racing to fix.
Anchor phrases narrow what's allowed to move
"Keep everything else" is the weak anchor — it's a wish, not a boundary. The strong version names what stays: "keep her face, expression, and pose identical" assigns specific elements to the freeze zone.
The pattern: name the change, then name the anchors. "Change the outfit to a black gown — keep face, hair, pose, and lighting identical." Every named anchor is a region the model has less license to repaint.
Chaining beats stacking: one change per pass. Three small anchored edits hold identity far better than one instruction asking to move wardrobe, light, and background together.
When drift is the feature
The same mechanism is desirable when you want transformation: photo-to-anime, photo-to-painting, heavy restyles. There you loosen the anchors on purpose — "redraw this scene in anime style" instead of "keep everything identical."
The dial isn't broken — it's just a dial. Knowing which direction to turn it is the skill: anchor hard for edits, release for restyles. The restyle playbook covers the release side; the character workflow covers the anchor side.
Prompts that work
"Change the background to a rainy street — keep her face, outfit, and pose identical." — swap with three named anchors.
"Relight to golden hour, keep everything else unchanged." — light pass, identity frozen.
"Same photo, now night — only the light changes, face and composition untouched." — time-of-day edit with the scope stated.
The honest limits
Anchors shrink drift, they don't erase it — extreme structural changes (hard pose contortions, full medium jumps) will reshape edges no matter how many anchors you name.
If an edit keeps drifting, shrink the ask. "Warmer light" holds tighter than "golden hour cinematic masterpiece" — scope creep in the prompt becomes scope creep in the render.