Most image model launches are about making pictures from nothing. Microsoft's newest one does the opposite. MAI Image 2.7, released in preview on October 27, only edits: you give it a photo and an instruction, and it gives you back the changed photo. Text-to-image is "coming soon", in Microsoft's words (Microsoft AI).

It got attention fast. Artificial Analysis's post announcing that the model had taken second place on its editing leaderboard passed 700,000 views within two days (X). US Google searches for "ai image editing" are also running at close to double their three-month average this week (Google Trends).

This post covers what the model is, how to read its rankings, and four edits from Microsoft's own list of strengths. We ran all four on models you can use today, with before and after images.

What MAI Image 2.7 is

From Microsoft AI's announcement and BigGo Finance's summary:

  • Editing only, for now. The input is an existing image plus a text prompt. It follows MAI-Image-2.6, which came out in August (Artificial Analysis).
  • What it is built for. Restyling or relighting a whole scene, changing text and symbols so they still look printed on the object, restoring and enhancing old assets, and edits that need an understanding of how objects, light and surroundings fit together. Microsoft's examples include moving a product into a new setting and changing the lighting and mood.
  • Where to try it. Arena, Artificial Analysis and the MAI Playground. Microsoft has not announced an API or pricing, and the weights are not open.

How to read the rankings

The headline claims come from two different leaderboards, which is why reports don't always agree:

  • Artificial Analysis (AA-Image-Editing v2.1): No. 2 with 1,193 Elo. Only OpenAI's GPT Image 2.5 Sunburst (max) is ahead, and it beats GPT Image 2.5 Flare (BigGo, Microsoft AI).
  • Arena Edit: No. 3, which makes Microsoft AI the No. 2 lab there (Microsoft AI).
  • Against its predecessor: it improves on 2.6 in all 10 use cases and all 7 editing actions that Artificial Analysis tests. It reaches the top tier in four categories: enhancement and restoration, reasoning-based edits, scene and style, and text or symbol editing. Its biggest lead is in scene and style, 24 Elo ahead of the next model.

Two caveats. These are preference votes on a set of test edits, not a measure of your own photos. And Elo gaps of a few dozen points are small, so the top three or four editing models are close enough that how you write the prompt still matters more than which one you pick.

Four edits worth trying today

We started with two plain product photos, both made with GPT Image 2 on EasyImagen: a ceramic coffee dripper on grey and a kraft coffee bag with a printed label. Then we edited them with Nano Banana and GPT Image 2, passing each photo in as the reference image.

A matte sage-green ceramic pour-over coffee dripper on a matching saucer, flat studio lighting on a light grey background

A kraft paper coffee bag on a light oak shelf with a cream label reading MORNING ROAST, Single origin · Medium, 250 g, under a line drawing of a sun

1. Move a product into a new setting (Nano Banana)

This is the edit online shops need most, and the easiest one to get wrong. The product has to stay exactly the same while the light around it changes.

Move this exact coffee dripper and saucer into a new setting: a sunlit wooden café counter in the morning, with a glass carafe and a paper bag of beans softly blurred in the background. Keep the dripper's shape, ribbed pattern, sage-green matte glaze and proportions exactly the same. Match the new light: warm sun from the right, a natural soft shadow on the wood, a faint reflection under the saucer. Photorealistic, no text, no logos.

The same sage-green dripper on a sunlit rustic wooden counter, with a glass carafe of coffee and a kraft bag of beans blurred behind it

The ribs, handle and glaze survived, and the shadow sits naturally on the wood. Naming the parts that make the product recognizable ("ribbed pattern, sage-green matte glaze") is what kept them. A vague "keep the product the same" gives the model too much room.

2. Change printed text and symbols (GPT Image 2)

Microsoft lists text and symbol editing as one of MAI Image 2.7's top categories. It is hard because the new words have to match the font, ink and paper texture of the old ones.

Edit only the label text on this coffee bag. Change "MORNING ROAST" to "EVENING DECAF" and change the sun line drawing to a crescent moon line drawing in the same thin black line style. Keep "Single origin · Medium" and "250 g" exactly as they are. Keep the same serif typeface, size, letter spacing, ink color and label position, and leave the bag, shelf, plant, mug and lighting unchanged.

The same coffee bag, now labeled EVENING DECAF with a crescent moon line drawing, while the smaller lines, shelf, plant and mug are unchanged

The new title matches the old typeface, and the two lines we told it to keep are untouched. One thing slipped: a few short rays from the sun drawing stayed around the moon. The fix is one more sentence, such as "remove the rays".

3. Relight the scene (Nano Banana)

Relighting means changing where the light comes from without moving anything. Describe the light as a physical source, and say where it falls.

Relight this scene as late evening: the window light is gone, the room is lit by a single warm table lamp just out of frame on the left, deep blue dusk visible through the window, soft warm highlights on the kraft paper folds and a long gentle shadow behind the bag. Keep every object, its position and the label text exactly the same. Photorealistic, cozy, low-key lighting.

The coffee bag at dusk, lit by a warm table lamp on the left with blue twilight in the window, label text unchanged

The mood change is convincing, and the label text came through letter for letter. But we asked for the lamp out of frame, and the model drew it in anyway. If the frame must stay clean, ask for "warm light from the left" and don't mention the lamp at all.

4. Restyle the whole image (GPT Image 2)

Scene and style is where MAI Image 2.7 leads by the widest margin. The test is whether the composition stays put while the medium changes completely.

Restyle this product photo as a loose ink-and-watercolor illustration on textured cold-press paper: fine sketchy brown ink outlines, soft sage-green watercolor washes with visible pigment blooms, a pale grey wash for the shadow, white paper showing through the highlights. Keep the same composition, angle and proportions of the dripper and saucer. No text, no signature.

The dripper redrawn as an ink-and-watercolor illustration with brown outlines and sage-green washes on textured paper

The angle and proportions match the photo closely, and the result reads as a real painting, not a filter. Naming the paper, the ink color and how the highlights should look gets you much further than just "watercolor style".

A pattern for any edit prompt

All four prompts follow the same three steps, and they carry over to MAI Image 2.7 or any other editing model:

  1. Say the one change, specifically ("change X to Y", "move this into Z").
  2. List what must stay, naming the features that matter: shape, pattern, typeface, position. The model protects what you name.
  3. Describe light physically: where it comes from, its color, where the shadow falls. This is what makes a pasted-in product look like it belongs there.

Then check the result for small misses, like the stray rays or the lamp above. Editing models are good enough now that the last 5% is usually one extra sentence.

Where this leaves you

If you want to try MAI Image 2.7 itself, the MAI Playground and Arena are the only options until Microsoft opens an API. For product and marketing images today, the prompts above work as written with GPT Image 2 and Nano Banana. For realistic food shots, see our guide to accurate AI menu photos. For seasonal designs, the Christmas card generator is a good place to start.