Meta AI vs Apple Intelligence: one Diwali kurta, one gym selfie

Same selfie, same Diwali kurta. Meta AI redraws you. Apple Intelligence dresses you.

Same photo: me in a white tee at the gym. Same words to both: "premium creamy Diwali kurta with gold embroidery."

• Meta AI in WhatsApp got the whole photo and sent back a new one.
• Apple Intelligence, through Image Playground inside my app Wearover, got only the part of the photo below my chin and drew a kurta for it. Wearover put that kurta back onto my original photo.

Watch the 13-second comparison (MP4)

What each one got right

Meta AI makes a lovely kurta. It's knee length, with heavy gold work on the collar, the placket and the cuffs. Because it redraws the whole picture, it can change how long the garment is.

The cost is that it redraws you too. Look closely and the face is very nearly mine, but it isn't my photo any more. And the whole photo, face included, went to Meta's servers to get there.

Apple Intelligence in Wearover drew cream silk with a gold-embroidered collar, placket and cuffs, and small gold motifs across the body. The face, hair and background are the original pixels. What left the phone was a crop of the clothes, sent to Apple's Private Cloud Compute.

The trade is length. Wearover only redraws the area the old top covered, so this kurta ends at the hip, not the knee. That's the next thing to fix.

How it works, in pseudo-code

// 1. Find the person and their clothes on the phone (Vision).
let person = try await vision.personMask(in: photo)
let top    = try await vision.clothingRegion(.top, in: photo, below: person.chin)

// 2. Read the words into a fixed shape (Foundation Models, @Generable).
@Generable struct Look {
    var colour: Colour        // cream
    var fabric: Fabric        // silk
    var style: Style          // festive  ← "Diwali" means festive
    var detail: Detail        // goldEmbroidery
}

// 3. Send Image Playground only the clothing crop, never the face.
let crop    = photo.cropped(to: top.box)          // starts below the chin
let drawing = await imagePlayground(crop, concepts: look.concepts)   // you tap ✓

// 4. Put the drawing back inside the old top's outline. Face and background untouched.
let result = photo.replacing(top.mask, with: drawing)

Two bugs this test turned up are already fixed in the next build:

• An older build added the default Navy swatch to my words. Image Playground got "navy top, creamy kurta" and drew a kurta that was half navy, half cream.
• "Kurta with gold embroidery" was split at "with", so the embroidery never reached Image Playground. Details like embroidery, zari and mirror work are now a fixed list the model picks from. They show up as a chip on the card, and Image Playground gets them as their own instruction.

Halloween, and tattoos

Same photo, one more ask: "Halloween Joker makeup." This one is different on purpose. A makeover has to change your face, so Wearover asks first ("Send your face to Image Playground?") and only then sends a crop of your face. Image Playground paints the makeup onto you, and Wearover blends it back into the photo with a soft edge. Undo puts your original face back, and Settings can turn makeovers off.

Tattoos work like the clothes: Image Playground draws the design from words ("delicate swallow in flight, black line-art"), and Wearover places it on the arm it found in your photo.

Sharing it

A look is only fun once someone sees it, so sharing got simpler. The share screen shows one big preview of exactly the picture you'll send, three thumbnails to pick the format, and one Share button:

• Look: the new look on its own.
• Before · after: side by side, or half and half in one photo.
• Story: a tall 9:16 card for Instagram and WhatsApp stories. The new look fills it, the original sits small in the corner, and the bottom line says where to get the app.

Every share carries the App Store link in its message.

Image Playground, under the hood

Image Playground is Apple's image generator. An app can open it as a system sheet, where you tap ✓ to accept a drawing, or ask for images in code. These are the parts of the API Wearover leans on:

• Concepts: short text pieces ("top in cream", "made of silk", "with gold embroidery"), plus an optional starting image. Wearover sends the clothing crop as that image, never the face unless you allow a makeover.
• Creation strategy: editExisting asks the model to stay close to the starting image and change only what the words ask for. That's what keeps your pose and your body the same.
• Size: sizeSpecification asks for a picture the shape of the crop, so it fits back without stretching.
• Personalization off: the model uses only the crop and the words, not your Photos library.
• Styles: a photo-like default, plus Animation, Illustration and Sketch. Clothes default to the photo look so they fit your photo. Tattoos and masks open as Illustration, and Sketch makes a good fine-line tattoo.

Which wins?

For a picture you'll post as you, Apple Intelligence through Wearover wins: your face is your face, your photo stays on your phone, and when you want your face changed, it asks first. For pure garment drama, Meta's longer kurta is ahead today, and that's the gap I'm closing next.

Wearover is on the App Store for iPhone and Mac.