Say What You See

Say What You See hands you an AI-made image and asks for a description precise enough to generate a convincing cousin. Three tries, a short text box, and rising match targets turn ordinary details like texture, lighting, medium, and composition into a surprisingly stubborn little puzzle.

What is Say What You See?

Each round begins with an image that looks easy to describe until the character limit starts biting. You type the most useful visual clues you can fit, Google creates another image from those words, and the game scores how closely the result resembles the reference. Passing thresholds climb as the levels advance, so broad labels quickly give way to questions such as whether the scene is a photograph or painting, where the subject sits, what the light is doing, and which materials or textures matter. Tips nudge you toward that vocabulary, while the three-attempt limit keeps every edit consequential. The clever twist is that the game grades observation through generation: you are not guessing a secret sentence so much as learning which details survive the trip from picture to prompt and back again.

What you can do there

• Describe a reference image in a short prompt
• Compare the generated image with the reference
• Refine the description using feedback

Why we picked it

It turns prompt writing into a compact visual parlor game instead of a blank text box. The similarity score is blunt, but it creates useful pressure to notice composition and medium rather than pile on decorative adjectives. A missed threshold can be irritating, yet the mismatch between the two images is often the most instructive part.

How to get the most from it

Name the medium and main subject first, then spend the remaining words on composition, lighting, color, and one distinctive texture. If the first result is wrong, change the detail most likely to alter the whole scene rather than rewriting every adjective. Later levels reward ruthless word choice.

Good to know

The game uses Google-generated imagery and expects English descriptions of up to 120 characters. Each picture allows three attempts, match targets rise after the first level, and a generated result may take several seconds. The underlying image model and current regional availability are not clearly identified.

Who made it, and when?

Jack Wild is the credited creator or organization. The earliest supported launch date we found is November 16, 2023.