← All benchmarks

Three Spotlights, One Wall

Red, green and blue spotlights overlap on a white wall. The prompt gives the arrangement and the three colours, and stops there. What colour each overlap turns is fixed by one rule the model is never told, and you can check every one of them yourself without leaving the picture.

6 models compared · same input · same prompt

Run it on your idea

Run your own prompt across all 6 models

Don’t just compare ours — get every model’s take on your idea, side by side. Full-res, one click. No juggling 6 accounts.

Input & prompt

Prompt template

A photograph of a white wall in a darkened room. Three round spotlights of the same size are projected onto the wall, one red, one green and one blue, arranged in a triangle so that each circle overlaps the other two, with a small area at the centre falling inside all three.

Before you scroll

What the correct answer looks like

Every patch of wall shows the sum of the lamps that reach it. Light adds. It does not mix like paint.

Overlapping lamps are the rare image-generation test with one right answer you can compute rather than argue about. Both renders below were produced by adding the lamp colours together — no artistic licence, no grading. Use them as the key when you scroll down to the model results.

Red, green and blue spotlights overlapping on a white wall. Red and green overlap as yellow, green and blue as cyan, red and blue as magenta, and the centre where all three meet is white.

Red, green, blue lamps

The version almost everyone has seen before. Pairs make yellow, cyan and magenta; all three make white.

Cyan, magenta and yellow spotlights overlapping on a white wall. Cyan and magenta overlap as blue, magenta and yellow as red, cyan and yellow as green, and the centre where all three meet is white.

Cyan, magenta, yellow lamps

The same rule with lamps nobody has memorised. Pairs make blue, red and green — and the centre is still white, because these are lamps, not inks.

The part almost nobody knows

Getting the colours right is the easy half, and it is the half that can be recalled rather than reasoned. The harder half is brightness: each time another lamp lands on a patch of wall, that patch gets brighter. The bars below are the actual measured brightness of each region in the renders above.

  • One lamp85
  • Two lamps overlapping170
  • All three — white255

Identical on both palettes, because the palette never mattered. This is also why a correct render looks dim: if three lamps must add up to white, each one can only be a third of the way there on its own.

What wrong looks like

All three of these show cyan, magenta and yellow lamps. Only one of them is light. The centre settles it on its own.

Correct additive render of cyan, magenta and yellow spotlights, with a white centre and every overlap brighter than the circles around it.

Correct — light adds

White centre, and every overlap brighter than the circles around it.

Incorrect render of cyan, magenta and yellow spotlights mixed subtractively like ink, where all three overlap the centre goes black and disappears into the dark wall.

Wrong — mixed like paint

The centre goes black and vanishes into the dark wall. This is ink on paper: it gets darker where more light lands, which is backwards.

Incorrect render of cyan, magenta and yellow circles averaged rather than added, with a flat grey centre.

Wrong — averaged, not added

Grey centre. The hues are right, so this one survives a glance — which is exactly why brightness is the test.

Every region in these figures is arithmetic on the lamp colours, not an artist's impression.

Open the full reference figure →

ChatGPT (GPT Image 2)

Try this →
openai/gpt-image-2

Nano Banana 2 (Gemini)

Try this →
google/nano-banana-2

Qwen Image 2.0 Pro

Try this →
alibaba/qwen-image-2-0-pro

Seedream 5 Lite

Try this →
bytedance/seedream-5-0-lite

Flux Kontext Pro

Try this →
flux/kontext-pro/text-to-image

Nano Banana 2 Lite

Try this →
google/gemini-3.1-flash-lite-image

Want to try this with your own twist?

Run the same prompt with your own variables — it's free.

Try the prompt →

New benchmark every week

We put the latest models head to head on the same brief — no labels, you judge. Get each new benchmark in your inbox the morning it drops.