Directing Materials & Finishes

When you write marble countertop, you have named a category, not a surface. The model has seen marble that is white and grey-veined, marble that is green, marble polished to a mirror and marble honed to chalk-matte, marble in slabs and in mosaic. Faced with that spread it does what a diffusion model always does: it samples toward the statistical centre of everything captioned 'marble', which is rarely the specific stone in your mind.

The fix is not more adjectives thrown at the wall - beautiful luxurious premium marble adds taste words that the model reads as mood, not material, and they steer almost nothing. The fix is to replace vague praise with discriminating facts: which marble, cut how, finished how, at what scale. Each fact narrows the region of images the model draws from until only the surface you want is left.

Think of it as removing guesses. wooden floor leaves the model six open questions. wide-plank smoked oak, matte oiled, wire-brushed grain, warm greyed brown answers all six, and the render snaps into focus because there is nothing left to invent. Specificity is not decoration - it is control.

There is a deeper reason this matters for design work specifically. A material is never neutral - it carries weight, cost, maintenance, acoustics and a whole cultural register. Kota stone and Italian marble might both read as 'grey floor' to a careless prompt, but one says vernacular and hard-wearing and the other says luxury and fragile. When you let the model average them away, you are not just losing visual accuracy, you are losing the argument the material was making. Precise naming keeps the render arguing the same case your specification does, which is exactly why it belongs in a designer's hands and not left to the model's defaults.

Every convincing material prompt fills six slots, and you can run them as a checklist:

1. Material / species - the base substance, named as specifically as a supplier would: not stone but honed Kota limestone; not metal but blackened mild steel; not wood but rift-sawn white oak.

2. Finish - what was done to the surface: oiled, lacquered, honed, polished, brushed, sandblasted, limewashed, powder-coated, board-formed. Finish is the single most under-used word and the one that changes the image most.

3. Texture - the tactile grain: wire-brushed, riven, bush-hammered, hand-troweled, ribbed, fluted, exposed aggregate. Texture tells the model how light should catch the surface at grazing angles.

4. Reflectivity - how the surface returns light: matte, satin, semi-gloss, high-gloss, mirror - or, in physically-based terms, low roughness (shiny) to high roughness (flat). This is the descriptor that reads most like the PBR parameters a render engine uses.

5. Tone / colour - the hue and value, described relatively: warm greyed brown, cool bone white, greenish patina, deep oxblood. Relative tone survives the model's colour drift better than a hex code it cannot honour.

6. Format / scale - the module and joint: wide plank, 600mm square tile, running bond brick, book-matched slab, standing-seam. Format anchors the material at the right physical size so it does not render as a miniature or a texture swatch.

You will not always need all six, but running the list stops you shipping nice tiles when you meant matte porcelain, 600mm, warm grey, minimal grout.

Modern render engines describe every surface with a small set of physically-based rendering (PBR) channels: base colour (albedo), roughness (matte to glossy), metalness (is it a metal), normal / bump (fine surface relief), and sometimes specular and transmission (for glass). Diffusion models were never handed these sliders - but they were trained on millions of captions and product photos that use exactly this vocabulary, so the words act like soft dials.

In practice: say low roughness, high specular or simply high-gloss, sharp reflections and the surface returns crisp highlights. Say matte, high roughness, diffuse and the same material goes flat and velvety. Add brushed metal, anisotropic streaks and you get the directional smear of satin steel. Add subsurface scattering, translucent and marble or alabaster glows from within. You are not toggling a real PBR channel, but you are naming the appearance that channel produces, and the model has seen enough labelled examples to comply.

The lesson: when a material looks wrong, diagnose it in PBR terms. Too plasticky? Raise roughness, kill the specular. Too flat and dead? Add a touch of gloss and a light source that can reflect in it - because reflectivity is meaningless without something to reflect, which is why materials and lighting (the next lesson) must be prompted together.

This diagnostic habit is what turns frustration into control. Most 'the AI just cannot do materials' complaints are really missing PBR words: a glass that reads as frosted plastic needs transmission, sharp refractions, clear; a fabric that reads as cardboard needs soft, matte, subtle sheen, woven texture; a metal that reads as grey paint needs metalness, brushed anisotropic reflections. You do not need the physics, only the vocabulary, and once you have it you stop rerolling the dice and start correcting a named fault. That is the difference between hoping the next generation is better and knowing which word will fix it.

Some finishes resist description. The exact travertine your client fell in love with, a bespoke terrazzo mix, a fabric with a particular weave - you can circle them in words but never quite pin them. This is where image-based material transfer earns its keep. Tools like Midjourney's image prompts and style-reference (--sref), or the open-source IP-Adapter on Stable Diffusion, let you feed a photo of the material itself and have the model carry its colour, grain and finish onto your scene while the text prompt controls the geometry.

Used well this is transformative: you shoot the actual sample under soft light, feed it as a material reference, and the render inherits a real finish instead of an invented one. Used carelessly it drags the reference's lighting and composition along too, so keep material references tightly cropped to the surface and lower their weight if they start rewriting your scene.

A disciplined workflow: describe the material in the six-slot text first, get the geometry and light right, then add a cropped material reference at modest strength to lock the exact finish. Text sets the scene; the reference fine-tunes the surface. On DesignAI the same principle applies when you restyle a room - you are transferring an appearance, not authoring a spec, so always confirm the real finish against a physical sample before anyone commits money to it.

One caution keeps this honest. A material reference transfers appearance, not behaviour - the model copies the look of the travertine but knows nothing of its porosity, its need for sealing, or how it will yellow near a window. So the reference-image trick is brilliant for communicating a finish and dangerous as a substitute for a real sample and a supplier's data sheet. Combine the two habits - six-slot text for the surfaces you can name, cropped references for the ones you cannot - and you cover almost every material a project throws at you, while never mistaking a convincing render for a checked specification. That discipline is the whole point: the model is directing pixels, and you are directing the material.