AI Moodboards & Palettes

People use the words loosely, but they are not the same thing, and confusing them is why so many AI moodboards go nowhere. A moodboard is a field of inspiration - textures, rooms, objects, light - assembled to communicate a feeling. It is deliberately loose; its job is to align you and the client on a direction before anything is decided. AI is very good at producing these quickly.

A palette is a specification - a small, closed set of colours and materials, each with a name, a role and an approximate share of the room. 'Warm off-white lime plaster (dominant, ~60%), oak and cane (secondary, ~30%), sage and brushed brass (accent, ~10%)' is a palette: someone could source against it. A palette is what actually drives a scheme, a restyle prompt, or a purchase order.

The workflow that works is: generate a moodboard, then _extract_ a palette from it. The moodboard is where you explore; the palette is where you commit. AI accelerates the first step dramatically and assists the second, but the act of assigning roles and shares - deciding which colour dominates and which is a whisper - is a design judgement you make, not one the model makes for you. Treat the moodboard as a question and the palette as your answer.

The naive way to build an AI moodboard - type a different prompt for each tile - produces exactly the incoherent wall described above, because every generation is an independent roll of the dice. Coherence has to be engineered, and there are three levers.

The strongest is a style reference: one image that anchors the look, which the tool applies across every tile. In Midjourney this is the style-reference feature (--sref); in the Stable Diffusion world it is an IP-Adapter, which conditions generation on a reference image's style rather than its content. Feed one anchor image - a room, a material, even a painting whose palette you love - and every tile inherits its mood. This single move does more for consistency than any amount of careful prompting.

The second lever is a fixed material-and-colour string repeated verbatim in every prompt: the same 'oak, lime plaster, cane, sage, brass' clause each time, changing only the subject ('a reading corner', 'a kitchen', 'a bedside'). Because the vocabulary is identical, the tiles share a family even when the rooms differ. The third lever is the seed: reusing one seed gives a set of images shared underlying structure, useful when you want variations that feel like siblings. Stack all three - one style reference, one material string, one seed family - and nine tiles read as one scheme instead of nine moods. That is a moodboard a client can trust.

Once you have a coherent moodboard, reduce it. Pull out five or six colours and materials, then give each one a job - because a palette without roles is just a colour list, and a colour list does not tell anyone how to build a room.

The old designer's heuristic that keeps a scheme calm is 60-30-10: roughly sixty percent a dominant, quiet field colour (usually walls and large surfaces), thirty percent a secondary that carries the material character (furniture, timber, larger textiles), and ten percent an accent that provides the spark (a cushion, a metal, a plant, one bold object). The numbers are not sacred, but the idea is: a room needs one clear lead, one strong support, and a small jolt - not five equal shouts. When you assign your extracted colours to these three tiers, a pretty grid becomes a specifiable scheme.

AI helps here in two honest ways. You can ask a vision-capable model to read your moodboard and propose the dominant, secondary and accent colours as named swatches with hex values - a fast first pass you then correct. And you can generate the palette back as a clean swatch strip for the client deck. What AI should not do is make the role decision for you: choosing which colour leads is where taste and the room's light and function come in. Extract with the machine; decide with your eye. The output of this step is the exact material string you will paste into a restyle prompt (6.1) and hand to a maker.

A single palette is easy; a project of moodboards - one per room, all clearly belonging to the same home - is where consistency is really tested. The failure mode is subtle drift: each room's board is fine alone, but side by side the whites don't match and the timber warms and cools from room to room. The fix is discipline, not luck.

Lock the shared variables and vary only the room. One master style reference for the whole project; one canonical material-and-colour string that every prompt inherits; a small, reused seed family. Change the subject slot and nothing else. Under Stable Diffusion, a tiny LoRA trained on your chosen material world can bake the palette into the model itself, so every generation is already in the family before you type a word - the studio-grade version of a fixed string. This is the same 'engineer the coherence' principle scaled from nine tiles to a whole deck.

And keep a palette source of truth outside the images: a small swatch card with names, hex values, materials and their 60-30-10 roles, written down once and referenced everywhere. Images drift; a written palette does not. When a new board wanders, you correct it back to the card rather than negotiating with the model. Tools like DesignAI, tuned for interiors, can hold this consistency for you across a project so you spend judgement on the scheme rather than plumbing on the coherence. The point stands whatever the tool: a coherent deck is decided once and enforced everywhere.

It is worth being explicit about why the reduction matters so much, because it is what connects this lesson to the rest of the module. A palette is not the end of a moodboard exercise; it is the input to every downstream step. The five-swatch string you extract is the exact material clause you paste into a 6.1 restyle prompt, so the render finally matches the concept. It is the vocabulary you hand the furniture and textile slots in 6.3, so the objects sit in the scheme rather than fighting it. And it is the finishes language a maker or a supplier can quote against.

This is the difference between a moodboard that decorates a pitch deck and one that drives a project. When the palette is written down with roles and shares, everyone downstream - the render, the schedule, the client, the vendor - is working from the same small, closed set of decisions. When it is left as a wall of images, each of those parties re-interprets it, and the scheme quietly diverges. The written palette is a coordination tool disguised as a colour exercise.

There is also a discipline in stopping. A moodboard can be explored forever - there is always another gorgeous tile to generate - and AI makes that temptation infinite. The palette is where you commit, and committing is a design act. Generate widely to explore the question; reduce ruthlessly to answer it; then hold the answer steady while you build. That rhythm - diverge, reduce, hold - is the professional habit underneath this whole lesson, and the palette card is the artefact that proves you have done it.