One thing I've noticed with image-generating algorithms is that the more of something they have to put in an image, the worse it is.

I first noticed this with the kitten-generating variant of StyleGAN, which often does okay on one cat:

A portrait of a black and white cat, looking at the camera with a shocked or possibly sarcastic open-mouthed expression
alternative for shocked_pikachu.png

but is terrible at a crowd of kittens.

One mostly normal-looking kitten surrounding by at least four others who are diffuse blobs of ears and eyeballs

A few years later, Dall-E 2 is a LOT more coherent despite having a larger job to do. But it's still susceptible to the kitten effect.

One kitten:

A single kitten in a basket. It looks almost photorealistic, though one eye is noticeably bigger than the other, and its whiskers are weird.
"A cute kitten in a basket"

Two kittens:

Two black and white kittens looking off to the left. Both have the correct number of eyes and ears and noses, but the shape of the head isn't quite right, and the eyes are blobby and unevenly positioned.
"Two cute kittens in a basket"

Ten kittens:

A basket of at least six kittens, who between them have about five complete faces. One kitten has too many toes on its paw and the other faces look a bit jumbled.
"Ten cute kittens in a basket". Note: the kittens are all happy and doing great because when you are a virtual kitten it doesn't matter if you have extra earholes in the middle of your face.

A vast herd of cute kittens thundering across the plain:

About twelve kittens, some of which are quite indistinct. One of them noticeably has an extra face on its belly.
"A vast herd of cute kittens thundering across the plain" These virtual kittens are also all healthy and happy. The one with extra eyes in its stomach is great at pouncing.

Similarly, there's a huge difference in quality between a single dog and a herd of them.

A husky dog running across a golden field. Its left paw is a bit blobby, but the rest of the dog is more or less in the right place and proportions.
A dog running across a field, matte painting
There is one dog with an identifiable face and tail and about three legs. The rest are increasingly abstract smears of fur and legs.
Vast herd of millions of dogs thundering across the plain, matte painting. They may look a bit odd, but they're all very good dogs.

And it's not just animals, of course. Here's the kitten effect played out in lucky charms marshmallows:

A single lucky charms marshmallow piece on a plate:

Each plate holds a single colorful marshmallow, some of which have rainbow stripes or swirls. The most interesting marshmallow shape is a star on toast. They're all pretty realistic looking though.

a single lucky charms marshmallow piece on a plate, labeled with the name of the shape:

The marshmallows, shaped like logs or clouds, are labeled Misily, Mhille, oLamum, Lamijal, Maaralla, and Masshaw

A labeled list of lucky charms marshmallows:

Handwritten chart with pastel-colored mini marshmallows scattered across it. Title reads "Charmus Macks", and the marshmallows are labeled "chabys", "Lamits", Mny Puuiks", and Lumaitrys.
Piles of mini marshmallows in bright colors with pastel splotches on them. The labels are pretty much illegible, although one reads something like Gonerealf

It was not obvious to me that this should be so. If it can make one kitten, why can't it make ten of equal quality?

My theory is this is may actually be not a numbers thing but a size thing. When I asked Dalle-2 to generate a kitten that takes up less of an image, the kitten gets way, way worse.

A 3d rendered-style room leading into a hallway and a mysterious shadowed far room. On the shiny wooden floor is a catlike shape, although it's smeary and indistinct.
"A kitten sitting at the far end of a large room".

It's not out of pixels to make the detail with - note the sharpness of the wood grain. But at this scale it has run out of something nonetheless.

It does make DALL-E 2's eye test charts particularly mean.

An eye chart with the top two rows looking normal, and the lettes in the lower row increasingly broken up or festooned with extra loops and hatch marks. By the last two rows, the letters aren't English characters anymore, though their shapes are still crisp.

Bonus post! In which I get DALL-E 2 to look increasingly closely at a giraffe.

Subscribe now