Google Labs Whisk Reviewed: Whisk It Up & Start Again

đź’ˇ
TLDR: Google Labs Whisk combines multiple images or text prompts to generate new images. While interesting, its unpredictable results offer little advantage over single, well-crafted prompts.

They should have called it Blender, but that name was already taken. As the name suggests, Google Labs Whisk is an interesting experiment that takes several different elements and whisks them together to form something new. So you add a subject (such as a photo of my cat Phoebe), a location such as a prehistoric forest, and a style, such as 1920’s animation. These can be a photo or a text description, which Google then uses to generate an image. After you’ve added the elements, it ponders for a bit and spits out a result.

USEFUL?

It’s fun but not very useful. The final result it creates by whisking the elements together is pretty predictable: once you understand the process, it would be quicker to develop the prompt yourself.

IN USE

I found that, like a lot of Google Labs experiments, it doesn’t always work that well. I started from scratch: I uploaded my photo, created an image of a prehistoric photo from a short description, and generated a realistic Betty Boop-style animation frame. However, the actual whisking didn’t work: I never got the final image.

Google Whisk in action: I uploaded the cat photo at the top left and created the next two images from a text prompt.

Okay, let’s try again, using one of the presets to create a push toy from another photo of my cat: relies, use the plushie preset and load the cat picture. After some time, it comes back with a plushie…of a young girl.

Plushie + Cat = Small Girls Doll Plushie?

Perhaps the Alpaca cat tree my cat was sitting on confused it? Let’s try again, with a simpler photo:

Much better. Let’s try tweaking the result so it looks more like my cat.

Click on refine and you can give more specific details. After a few rounds, it looked somewhat closer to the real thing.

However, no matter how much I argued with it, I couldn’t get rid of the stripes. Suggesting the refinement of “tortie pattern coat” turned her into a tiger for some reason.

Let’s try something else, using the sticker preset to create a sticker of Sundar Pichai, the CEO of Google that I can put on the back of my MacBook Pro.

Unfortunately, it really doesn’t look that much like him: the skin tone is wrong, the face shape is wrong, and the beard shape and color are all wrong. One of the created images also has a fake @handle on the outside, for no obvious reason, presumably from the source material.

A Stickie of Sundar Pichai?


Trying the from scratch version again, with a cat photo, a prehistoric forest and Betty boop-style animation, and it worked this time.

Our result is a decent image of a cat that looks a bit like Phoebe lounging in a prehistoric forest while a dinosaur wanders in the background—but still with the stripes! If you click on the image, you can see the full prompt that Google built to create it.

You can tweak this prompt to get closer to the results you want. For example, I removed the hexagonal pattern from the ground, made the dinosaur bipedal, and removed the odd object from the bottom right corner.

I am not sure if the cat is more concerned about being eaten or having two tails

Okay, a few steps forward (the ground and the dinosaur), but the cat now has two tails and looks rather concerned about the whole situation. Perhaps not surprising, but the cat still has stripes. Note to Google: Tortoiseshell cats don’t have stripes!

Have you tried Whisk? Let us know in the comments and post your images, even if they don’t have stripes.