AI image editing: seven kinds of edit, and how to ask for each

What this page can and cannot show

• Inpainting

  Replace or remove one region and leave the rest alone.

  Good for

  Removing a stranger or a blemish, swapping a small object.

  Usually fails by

  Changes leak outside the region, texture or light does not match, faces drift.

  How to ask

  Say what to remove or replace and where, then list what must stay unchanged.

  Outside ZeroTwo

  Dedicated inpainting models with a brush mask, for example FLUX.1 Fill.

• Outpainting

  Extend the canvas past the original frame.

  Good for

  A new aspect ratio, more headroom, a banner from a square.

  Usually fails by

  Visible seams, repeated textures, invented subjects.

  How to ask

  Ask the backdrop to continue, not for new subjects. Name the target ratio.

  Outside ZeroTwo

  Editors with an expand-canvas tool.

• Cut-out edits

  Isolate a subject or region, then recolour, replace or extract it.

  Good for

  Background swaps, recolouring one object, sticker cut-outs.

  Usually fails by

  Halos around hair and fur, rough edges in clutter.

  How to ask

  Describe the subject to keep and the new background. Match light direction and warmth.

  Outside ZeroTwo

  Segmentation models such as Segment Anything 2 and dedicated background removers.

• Style transfer

  Re-render the same content in another look.

  Good for

  Photo to watercolour, anime or pointillism.

  Usually fails by

  Structure drifts and detail is lost. A whole-image restyle is a re-render, not a filter.

  How to ask

  Name the style and say what to preserve: pose, composition, expression.

  Outside ZeroTwo

  Style-reference features in other generators.

• Upscaling and restoration

  Raise resolution and recover or invent fine detail.

  Good for

  Print enlargements, old scans.

  Usually fails by

  Invented detail, plastic skin, altered faces.

  How to ask

  Ask for faithful enlargement with no new elements, then check faces at 100%.

  Outside ZeroTwo

  Dedicated upscalers such as Topaz Gigapixel or SUPIR. A general edit model can re-render detail, so use one of these when fidelity matters.

• Generative insert

  Add an object that was not in the photo and make it belong.

  Good for

  A product in a lifestyle scene, staging, extra props.

  Usually fails by

  Wrong light direction, wrong scale, missing shadow or reflection.

  How to ask

  Say where it goes, how big it is next to something in the scene, and where the light comes from.

  Outside ZeroTwo

  Workflows built around a reference-image adapter.

• Identity-preserving re-render

  Change pose, outfit, lighting or setting while the subject stays recognisably the same.

  Good for

  Same person in a new outfit, same product in new scenes.

  Usually fails by

  Subtle face drift, changed gaze, altered clothing details.

  How to ask

  List the features that must not change (face, hair, eye colour) before saying what should.

  Outside ZeroTwo

  Face-consistency tools such as PuLID and InstantID.

A crayon drawing, made lifelike

Before

After

Kept

The round ears, the smile, the upright front-facing pose and the red-orange colour.

Changed

Body proportions and the scribbled torso became fur. Whiskers and a curled tail appeared, and the lined paper was replaced by a plain cream ground.

A sunset photo, restyled

Before

After

Kept

The sunset palette, the palm trees, the pool reflection and the loungers.

Changed

The framing: the output is portrait while the source is near-square, and the chairs, pool edge and trees sit in different places. A stipple texture covers everything.

The lesson from the second pair: a whole-image restyle is a re-render, not a filter. When the composition must stay exact, ask for a narrower edit and list what to preserve. More example prompts live on the image generation feature page.

The seven edit-capable models in ZeroTwo's list matched to their closest Artificial Analysis image-editing entry, checked 5 October 2026: maker, closest board entry, rank out of 80, Elo and a note on how sure the match is.

Edit-capable in ZeroTwo	Maker	Closest board entry	Rank of 80	Elo (95% interval)	Note
GPT Image 1.5	OpenAI	GPT Image 1.5 (high)	12	1102 ± 8	Name match.
Nano Banana Pro	Google	Nano Banana Pro (Gemini 3 Pro Image)	13	1099 ± 8	Name match. Ranks 12 and 13 overlap, so treat them as a tie.
Flux Pro 2	Black Forest Labs, via FAL	FLUX.2 [pro]	41	1007 ± 7	Closest entry. We cannot confirm which FLUX.2 variant ZeroTwo calls Flux Pro 2.
Nano Banana	Google	Nano Banana (Gemini 2.5 Flash Image)	50	986 ± 9	Name match.
GPT Image 1	OpenAI	GPT Image 1 (high)	53	979 ± 7	Name match, listed at its high setting.
GPT Image 1 Mini	OpenAI	GPT Image 1 Mini (medium)	66	923 ± 7	Name match, listed at its medium setting.
Qwen Edit	Alibaba	Qwen Image Edit	68	918 ± 7	Closest entry. The board lists several Qwen edit versions; we cannot confirm which one ZeroTwo uses.

Source: Artificial Analysis, AA-Image-Editing v2.0, read 5 October 2026. Matches marked "closest" are our judgement, not Artificial Analysis's.

• Remove a stranger from a holiday photo

  Inpainting

  First moveName the person and where they stand, then list what must stay unchanged.

  If it goes wrongOne change per turn. If your tool offers a brush mask, use a tight, slightly soft one.

• Turn a square crop into a 16:9 banner

  Outpainting

  First moveAsk the existing backdrop to continue left and right; add no subjects.

  If it goes wrongExtend in smaller steps, or crop tighter first so there is less to invent.

• Swap a beach background for a studio backdrop

  Cut-out edit

  First moveSay to keep the person exactly as they are, then describe the new backdrop and its light.

  If it goes wrongCheck hair edges at 100%. Fine fur and flyaway hair are where halos appear.

• Make a photo look like a watercolour

  Style transfer

  First moveName the style and say to keep pose and composition exactly.

  If it goes wrongIf the layout still drifts, accept it as a re-render and pick the version closest to the original.

• Enlarge a 1024px headshot for print

  Upscaling

  First moveUse a dedicated upscaler when fidelity matters. Compare free upscalers.

  If it goes wrongCheck eyes, teeth and skin at 100%. Invented detail shows up there first.

• Put a sneaker on a marble countertop

  Generative insert

  First moveAttach a clear product photo, say where it sits, how big it is and where the light comes from.

  If it goes wrongName the shadow direction explicitly. A wrong shadow is the quickest giveaway.

• Same person, new outfit

  Identity-preserving

  First moveList face, jawline, eye colour and hairstyle as fixed, then describe only the clothing.

  If it goes wrongShorten the list of changes. Identity drifts more the more you ask for at once.

• Restore a faded family photo

  Restoration, then inpainting

  First moveChain two edits: restore tone and detail first, then fix scratches or tears.

  If it goes wrongKeep the original scan. Compare faces against it before you accept a result.

• Remove a person

  Inpainting

  Remove the person standing at the far left of the frame, next to the lamp. Fill the space with the surrounding background so it looks as if they were never there. Keep everything else unchanged: lighting, colour grade, framing and every other person. No text, no logos.

  Why: It says what to keep as well as what to remove, and it locates the region in words so it does not depend on a mask tool.

• Extend to 16:9

  Outpainting

  Extend this image to a 16:9 landscape format by continuing the existing studio backdrop to the left and right in the same grey gradient. Keep the subject exactly as it is, keep the depth of field and film grain consistent, and add no new subjects.

  Why: Extensions fail when you ask for new content. Restricting the request to backdrop continuation is safer.

• Photo to watercolour

  Style transfer

  Re-render this photo as a loose watercolour with visible paper texture and soft bleeding edges. Keep the pose, expression and composition exactly as they are.

  Why: The explicit keep-list guards against structural drift, which is what the pointillism example above shows.

• Same person, new outfit

  Identity-preserving

  Keep this person's face, jawline, eye colour and hairstyle identical. Change only the clothing to a tailored navy blazer over a white oxford shirt. Soft studio light from camera left.

  Why: Face drift comes from an under-specified keep-list, so name the features before the change.

• Place a product in a scene

  Generative insert

  Add the product from the attached reference image onto the marble countertop, centred. The light comes from the window at upper right, so cast its shadow to the lower left and add a faint reflection on the marble. Keep it about the size of a real shoe next to the glass.

  Why: Naming the light source and the scale gives the model what it needs to make the object belong.

• Swap the background

  Cut-out edit

  Keep the person exactly as they are and replace the background with a softly out-of-focus city street at golden hour. Match the light on the person to a warm late-afternoon colour and add a faint warm rim light on their left shoulder.

  Why: Matching light and colour temperature is what separates a believable composite from a cut-out.

1. 1

   Did you name the edit type first?

   If you are fighting a general tool with an edit it was not built for, say inpainting for an upscaling job, the fix is a different approach, not a longer prompt.

2. 2

   Is the request one change?

   Several changes in one turn compete. Do them one at a time and keep the version you like.

3. 3

   Did you say what to preserve?

   Identity, colour and composition drift when the keep-list is missing. Add 'keep the face', 'keep the colour grade' or 'keep the composition'.

4. 4

   Is the region clear?

   If your tool has a mask, make it tight and slightly soft-edged. If it does not, locate the region by position and neighbours, such as 'the person at the far left, next to the lamp'.

5. 5

   Is this a model limit?

   Try another of the seven edit-capable models. If it still fails, finish the last stretch in a layered editor; that is a normal workflow, not a defeat.

• Check at full size

  Edits that look fine as a thumbnail often show seams, hands or text errors at 100%.

• Disclose where it matters

  Where an audience would expect an unedited photo, such as news, evidence or a product listing, say the image was edited.

• Do not strip other people's marks

  Using an edit to remove a watermark or copyright notice from an image you do not own, in order to reuse it, can be unlawful in many places. This is general information, not legal advice.

• Best AI photo editor

  Compare photo-editing products by task.

• Best AI photo batch editor

  Bulk workflows for hundreds of photos.

• Best free AI upscaler

  Enlarge images faithfully for print and 4K.

• Best AI image generator

  A dated leaderboard snapshot and a job-by-job guide.

• Text-to-image prompt cookbook

  When you are starting without a source image.

• Free AI image generators

  Quotas, watermarks and licences for no-cost tiers.

• AI image to video

  Turn an edited image into a short clip.

• Image generation in ZeroTwo

  The feature page, with example prompts.

Broader questions: how ZeroTwo compares as a single workspace is covered in the best all-in-one AI platform guide, the chat side in ZeroTwo's AI chat, and the text models in AI models for text generation.

Upload a photo and try one edit from this page