Generating an AI image from a moodboard: the majority wins

You can generate from a moodboard by attaching its pictures as style references — up to 3 on the $8 plan, 10 on the $20 one. But a board is not averaged. We attached four references that disagreed, two warm and two cold, and the warm pair won outright: the cold pictures left no trace at all. A moodboard works when it already agrees with itself, and the tidying is your job, not the model's.

The experiment

Four style references, deliberately mixed: a cool blue temple lamp at dawn, a warm oak kitchen counter, a black-and-blue engraved cyber portrait, and a warm sunlit cottage interior. The prompt was plain — a wooden writing desk by a window with an open notebook and a cup.

Everything cold vanished. No blue, no black, no engraved line, nothing of the temple or the portrait. The result is entirely warm wood, cream linen and rough plaster — the two warm references, and nothing else. The stoneware cup even carried its two-tone glaze across from the kitchen picture, despite that reference being badged Style, which is supposed to ignore subjects.

What that means for building a board

• The majority wins; it does not blend. Two references against two produced a result with no trace of one pair. Do not expect a board to meet in the middle.
• So curate before you upload. Ten pictures you like is not a moodboard. Ten pictures that agree about light, palette and surface is.
• Three agreeing references beat ten mixed ones. Every image is another vote, and votes for different things cancel rather than combine.
• Strong subjects leak. A style reference with a very distinctive object in it can still push that object into your picture. If that happens, crop the reference down to its texture and light before uploading.

What a coherent board does instead

Here is the same technique with a board that agrees — three pictures of one wet blue temple dawn, against a prompt about a clay pot that mentioned no weather at all.

Building one, step by step

1. Lay your board out and throw away every picture that disagrees with the majority about light or palette. This is the whole job.
2. Crop any survivor whose subject is louder than its mood, so what remains is texture, colour and light.
3. Upload three to ten of them and badge each one Style, which tells the model to take palette, mood and rendering and ignore subject and composition.
4. Write your own subject plainly. Do not describe the mood as well — the board is doing that, and two descriptions fight.
5. Generate at 1K for 8 credits. References cost nothing extra.

The cheaper route, and when to prefer it

If your board is really about colour rather than surface, decode one picture from it for 3 credits, take the five hex codes from its VISUAL tab and name them in the prompt instead. That is more controllable than an upload, because you assign each colour to a part of the picture rather than hoping the average lands. Uploads are better when what you want is material — wet stone, worn brass, coarse linen — which no colour list carries.

FAQ

Can I generate an AI image from a moodboard?

Yes — upload the board's pictures as style references, three on the $8 plan and ten on the $20 one, and badge each Style so the model takes the palette, mood and rendering rather than the subjects.

Does the model blend all the references together?

No, and this surprises people. In our test two warm references and two cold ones produced a result with no trace of the cold pair at all. The majority wins; references do not meet in the middle.

How many pictures should a moodboard have?

Three that agree beat ten that do not. Every image is another vote, and votes for different things cancel out rather than combine. Curate before you upload.

Why did an object from my moodboard appear in my picture?

Because a strong subject can leak through even under the Style role. Ours carried a cup's two-tone glaze across. Crop the reference down to its texture and light before uploading if that happens.

Is there a more controllable alternative?

For colour, yes: decode one board picture for 3 credits and name its hex codes against parts of your image. Uploads remain better for material — wet stone, worn brass, coarse linen — which a colour list cannot carry.

Throw half your board away, then upload the rest in the Studio and badge them Style. What each reference role actually tells the model is here.