Speakasso › Guide
GuideHow to describe a painting so AI can draw it
Describing a picture well is a skill: you have a few seconds, a limited number of words and a listener — an AI — that takes everything literally. This guide is the method that scores highest in Speakasso. It also works for writing prompts for any AI image generator.
Why it’s harder than it looks
When we look at a famous painting, we recognise it and want to say its name. The AI can’t use the name — in the game the title, the artist and the key objects are forbidden words — so you have to rebuild the picture from scratch: what is there, where it is, what it looks like.
People also tend to describe the mood (“it feels calm and mysterious”) instead of what is visible. Mood alone gives the AI almost nothing to draw.
1. Start with the main subject
The subject is worth the most — 50 of 100 points in Speakasso. Say who or what is in the centre of attention, how many there are, and what they are doing.
- Who or what: a woman, two fishermen, a village, a bowl of fruit.
- Pose and gesture: sitting at a table, turning her head over her shoulder, rowing.
- Details that make it this picture: a blue and yellow turban, a pearl earring, a letter in her hands.
2. Say where things are
Layout and setting are worth 20 points. The AI places objects wherever it likes unless you tell it.
- Foreground and background: in front, behind, in the distance.
- Left, right, centre, top, bottom.
- The horizon: low or high, and what is on it — hills, a town, the sea.
3. Name the colours and the light
Colour and light bring another 20 points and are the quickest to say.
- Two or three dominant colours: deep blue and bright yellow, muted brown and gold.
- Time of day and light source: night with glowing stars, a window on the left, a sunset behind the hills.
- Contrast: a dark background with one brightly lit face.
4. Add the style last
Style and technique are worth 10 points: useful, but only after the rest. Say how it is painted, not who painted it.
- Thick swirling brushstrokes, smooth glossy oil, flat woodblock print, soft blurry dabs of colour.
- Never guess a style you can’t see — a wrong style changes the whole picture.
5. Describe instead of naming
Forbidden words cost 15 points each. Every forbidden word can be replaced by what it looks like: not “cypress” but “a tall dark tree shaped like a flame”; not “the moon” but “a bright crescent in the top right corner”. The AI draws from the look, not from the name.
Example: The Starry Night
A description that scores high without a single forbidden word:
“A night sky full of huge swirling spirals of wind and glowing stars, a bright yellow crescent in the top right. On the left in the foreground, a tall dark tree like a black flame rises to the top. Below, a small sleepy village with a church spire and lit windows, blue hills behind it. Deep blue and bright yellow, thick swirling brushstrokes.”
Subject, layout, colour and light, style — in about 20 seconds of speech.
The mistakes that cost the most
- Naming the painting or the artist — minus 15 each, and the AI still doesn’t see the picture.
- Only the mood — “beautiful, sad, dramatic” — with no objects.
- A long list of small details before the main subject is clear.
- Inventing things that aren’t there, including a style you are not sure of.
The same skill makes better AI prompts
Subject → layout → colour and light → style is also a reliable order for prompts in any AI image generator. Describe what is visible, say where it is, and leave names out: you get what you described, not what the model associates with a title.
Try it in a duel
Speakasso is coming to Google Play. Leave your email and we’ll tell you on launch day.