Ignore the Cozy Advice for Keycap Diorama Prompts

Macro photography, extreme close-up of a vintage cream keyboard keycap hollowed into a miniature rustic bedroom diorama, weathered ABS plastic with "hme" carved into the top edge (implying "home" via negative space), reclaimed barn wood plank walls with visible grain and oxidized nail holes, distressed oak floor, tiny handcrafted oak bed frame with rumpled oatmeal linen sheets and ticking stripe throw, textured pillows in sand and wheat tones, miniature brass table lamp with pleated shade as the ONLY light source, warm 2700K glow pooling on the right wall and fading to shadow on the left, no ambient fill, small frayed jute rug, single worn leather slipper under the bed, surrounding keycaps dissolving into creamy bokeh, f/2.8 shallow depth of field, chiaroscuro contrast between lit interior and dark keycap shell, earth palette only of raw umber, aged ivory, tobacco brown, antique brass, photorealistic micro-textures, 8k, shot on 100mm macro lens --ar 9:16 --style raw --s 250

The most repeated advice for keycap diorama prompts is to call the scene "cozy" and stack mood words on top: warm, inviting, homely, nostalgic. That advice reliably makes the output worse. Diffusion models do not resolve mood, they resolve light, geometry and material. A cluster of emotional adjectives collapses into one global color grade, which lifts shadows across the whole frame and flattens exactly the local contrast that makes a miniature read as miniature. The fix is not more atmosphere. It is naming one light source, giving it a temperature, and stating where its falloff ends.

The prompt box above does that. Instead of mood, it supplies a physical inventory: a brass table lamp with a pleated shade as the only source, 2700K, pooling on the right wall and fading to shadow on the left. Every clause is measurable. That single substitution — adjectives replaced by a lighting diagram — separates a diorama that looks lit from one that looks painted.

The Shell and the Negative-Space Carve

Start with the container, because everything else is scaled against it. "Vintage cream keyboard keycap" does more work than it appears. Cream, not white, prevents the model from rendering clean high-key plastic; it biases toward the slightly yellowed tone of aged ABS. Adding "weathered ABS plastic" locks the surface into matte micro-pitting rather than the glossy injection-molded finish that a bare "keycap" keyword usually produces.

Hollowing is the second half of the shell instruction. "Hollowed into a miniature rustic bedroom diorama" tells the model that the keycap is a cutaway box, not a solid object with a room painted onto it. That distinction controls whether the interior walls are lit by the lamp inside or by the ambient scene outside. Without the word hollowed, most models render the interior as a flat illustration sitting inside a letterform.

The carved "hme" is the most fragile element in the prompt. Letterforms are read as texture, and an incomplete word invites autocorrection to "home." The phrasing "carved into the top edge (implying 'home' via negative space)" counters this by describing the missing letter as an absence rather than a typo. Incised geometry also renders differently from printed legends: carved letters catch shadow on one edge, which reinforces the single-source lighting specified later.

Before moving on, check the shell. The top edge should show letterforms cut into the plastic with a shadowed inner edge, and the interior should be visibly recessed behind the front face. If the letters look printed and flat, the carve instruction is being ignored — shorten the word, keep "negative space" explicit, and repeat "weathered ABS" instead of "keyboard keycap."

One Lamp, 2700K, and a Stated Falloff

Two light sources destroy a miniature. With a lamp plus ambient fill, the model has no reason to darken anything, and the interior flattens into an evenly lit dollhouse. The prompt above avoids this with an absolute: the brass table lamp is the ONLY light source. Capitalized exclusivity is not decoration; it removes the fill light the model would otherwise add by default, and it is the clause that makes the chiaroscuro request physically coherent.

Temperature is the next lever. 2700K is not a synonym for warm — it is a specific point on the blackbody scale, roughly the color of a domestic incandescent bulb. Naming the number rather than the adjective keeps the highlight from drifting toward the orange-amber cast that "warm golden light" tends to produce, and it keeps the shadow side neutral rather than purple. The result reads as a real bulb instead of a color grade.

Direction matters as much as color. "Pooling on the right wall and fading to shadow on the left" gives the model a gradient with an origin and a terminus. That produces the lit-interior-versus-dark-shell contrast: the right wall sits near the bulb, the left wall falls off, and the outer keycap surface stays dark because no light reaches it. Without the directional clause, models often center the glow and lose the asymmetry.

Check this section by looking at the shadow side of the bed. It should be visible but clearly darker than the wall behind the lamp, with no second highlight on the left. If the left side is bright, the exclusivity clause is being diluted — remove any word like "ambient," "soft glow" or "fill" and test again. Parameters interact here: --style raw reduces the default aesthetic grade that would otherwise warm the whole frame, while a stylize value near 250 keeps the render literal instead of illustrative. The Midjourney documentation describes stylize as a measure of artistic interpretation, which is why high values soften hard lighting ratios.

Material Vocabulary at Miniature Scale

Scale is communicated by material, not by the word miniature. "Reclaimed barn wood plank walls with visible grain and oxidized nail holes" supplies two scales of detail — plank-level grain and millimeter-level nail corrosion — and the presence of both forces the camera to sit close. A single texture word such as "wood walls" produces one frequency of detail, which reads