ChatGPT Image Editor

No. The editor runs OpenAI's and Google's image models directly, and your first edits need no account here either.

It is not supposed to, and holding identity across an edit is exactly what the pairs above are demonstrating. Name what must stay the same in your sentence and it holds far more reliably.

Yes, but you will get a more predictable result changing one thing per pass. Two instructions in one sentence is where results start drifting.

Twenty credits at 1K, forty at 2K, eighty at 4K — shown before you run it, and refunded if the run fails.

That has its own page — the background remover returns a transparent PNG rather than a new scene.

Run the result through the image upscaler afterwards. Edit first, enlarge second.

A selection tool asks you to trace the boundary of the thing you want to change, which is slow on hair and impossible on soft edges like smoke or fur. Describing skips the boundary entirely: you name the region in words and the model works out where it is. The trade is control — a mask does exactly what you drew, and a sentence does what it understood — which is why naming what must NOT change matters as much as naming what should.

As many as you like, but each pass works from the previous result rather than from your original, so small errors compound. Three or four passes is where most people start seeing drift in a face. If you need a big change, it is usually cheaper and cleaner to write one instruction that covers it than to arrive there in five steps.

Anything where the instruction does not match the content. An edit prompt that names pose and expression will confuse a model given a product shot, because a bottle has neither — we measured that directly while building the examples on this page, and it produced reinterpretations rather than relocations. Match the wording to what is actually in the frame.