Two prompts, one word apart. “A wooden treasure chest with iron bands” produced a 41,000-triangle mesh with rounded lid, smooth transitions, and iron banding that was entirely painted into the albedo. “A low poly wooden treasure chest with iron bands” produced 3,600 triangles, flat-shaded facets, a lid built from four visible planes — and the same painted banding. One word changed the entire topology. The rest of the sentence changed nothing about the form at all.
This is the split most people never see. Text-to-3D runs geometry and surface as related but separate concerns, and roughly half of what a typical prompt says lands only on the texture. If you are trying to fix a silhouette by adding “rusty, weathered, highly detailed” you are writing instructions for the paint job and wondering why the shape does not change.
Which words move vertices
The useful test is whether a phrase describes something a modeller would build or something a texture artist would paint. “Panelled” is built. “Scratched” is painted. “Chamfered” is built. “Chrome” is painted. Words about construction, proportion, silhouette, thickness, symmetry and topology reach the mesh. Words about colour, wear, material name, finish and grime almost never do.
| Prompt phrase | Concrete geometric effect |
|---|---|
low poly |
Flat-shaded facets, hard normals, typically 2k–5k triangles; secondary detail collapses into primary forms |
faceted |
Planar polygonal surfaces with visible flat panels, like a cut gemstone; distinct from low poly in that density stays moderate |
hard surface |
Beveled panel edges, cut seams and panel lines as real geometry, mechanical joints, large planar regions with sharp transitions |
organic |
Continuous curvature, no sharp creases, denser and more evenly distributed triangles across the surface |
chamfered edges / beveled edges |
1–3 mm rounded edge loops that catch a specular highlight; adds roughly 15–25% to triangle count and dramatically improves how the model reads under any lighting |
sharp creases |
Near-zero-width edges; crisper silhouette, but edges disappear under soft lighting |
chunky proportions / thick |
Increased wall and limb thickness, small protrusions merged into the body, more forgiving silhouette at small screen sizes |
thin walled |
Attempts sub-millimetre shells; frequently produces holes and non-manifold edges. Use sparingly |
hollow, open interior |
Generates an interior shell with visible thickness at the rim rather than a solid plug |
solid closed form |
Watertight single shell, no interior cavity; best choice for anything going to 3D print |
symmetrical |
Mirrored halves with a clean centre seam; makes later retopology and rigging far easier |
asymmetric |
Different left and right halves; more interesting silhouette, harder to clean up |
flat bottom, sits flat on a surface |
Planar footprint at the base, no rocking; the origin ends up somewhere usable |
on a circular base / on a plinth |
Adds a discrete base disc, changes the bounding box, stabilises tall thin subjects that would otherwise generate broken supports |
modular kit piece, flat back |
Bounds the geometry to a box with one planar face; the back is culled flat instead of wrapped |
greebled / panelled surface detail |
Raised plates, vents and boxes as actual protruding geometry; typically doubles triangle density |
engraved / embossed relief |
Shallow real displacement in the geometry stage; note this frequently gets replaced by a normal map if you refine afterward |
battle damaged, dented, crumpled |
Genuine surface deformation — dents that change the silhouette |
scratched, rusted, worn paint |
Texture only. Zero effect on the mesh |
cylindrical body / boxy / tapered |
Forces the primary form toward that primitive; the most reliable way to control overall shape |
exaggerated stylized proportions |
Enlarged primary features, compressed secondary ones; reads correctly at 64 px on a UI card |
separate detachable parts |
Multiple disconnected shells in one file; useful for assemblies, breaks any pipeline expecting a single closed mesh |
fur, hair, feathers |
Produces a fuzzy amorphous blob. Model the body, add fur with cards or a shader |
transparent glass |
Geometry is usually fine, shading is not; ask for solid form and assign a transmissive material yourself |
Words that do nothing, and words that actively hurt
A large amount of prompt folklore migrated over from image generation and does not survive the trip. “Octane render”, “unreal engine 5”, “8k”, “trending on artstation”, “cinematic lighting”, “shot on 85 mm” — none of these describe form. At best they are ignored. At worst, the lighting terms bake specular highlights and shadow gradients into the albedo, which is exactly what you spend time removing later.
Triangle counts in the prompt do not work either. Writing “8000 triangles” gets you a model with whatever density the system was going to produce anyway; the number is not a parameter the text encoder can act on. Use the target polycount control in the interface instead, which actually drives the remesher.
Prompt length has a real ceiling. Between about 25 and 60 words is where the useful range sits. Past roughly 80 words the tail of the prompt measurably stops influencing the result, and you end up with a paragraph you believe is doing work that is not. If you have more to say than that, you have a job for two generations and a kitbash, not one long sentence.
Structure that works
Front-load the noun. The subject should be within the first four or five words, because everything after it is read as modifying that subject. “A rugged, heavily weathered, dark green military field radio” is worse than “A military field radio, rugged and heavily weathered, dark green” for exactly this reason.
Then build outward in this order:
- Subject. One concrete noun phrase. “A wall-mounted electrical fuse box.”
- Primary form. The overall shape and proportion. “Rectangular steel enclosure, taller than wide, flat back.”
- Construction detail. What a modeller would actually build. “Hinged front door with a recessed handle, three conduit stubs at the bottom, beveled edges.”
- Style and topology hint. “Hard surface, stylized, chunky proportions.”
- Surface, last and briefly. “Chipped grey paint, rust at the seams.”
Here is that prompt evolving across three iterations, which is roughly how long it should take:
v1: “electrical fuse box, industrial, rusty” — 28k triangles, a vaguely box-shaped lump, everything interesting painted on, back face wrapped and unusable for wall mounting.
v2: “A wall-mounted electrical fuse box, rectangular steel enclosure, flat back, hinged front door, industrial, rusty” — flat back solved, door now a separate plane, still soft and undefined at the edges.
v3: “A wall-mounted electrical fuse box. Rectangular steel enclosure, taller than wide, flat back for mounting. Hinged front door with recessed handle, three conduit stubs at the bottom, beveled edges. Hard surface, chunky proportions. Chipped grey paint.” — 9,400 triangles, readable panel breakup, conduits present as geometry, mountable.
Notice that “rusty” survived unchanged through all three versions and never once affected the mesh. It was never going to. The gains all came from form language.
The same pattern shows up whenever you A/B a single phrase against a fixed base prompt. These are typical results from one afternoon of testing on a “wooden treasure chest” base, all at default preview settings:
| Added phrase | Triangles | Silhouette change | Usable without cleanup? |
|---|---|---|---|
| (base prompt only) | 41,200 | — | No — dense, uneven density |
rusty, weathered, highly detailed |
40,800 | None measurable | No |
low poly |
3,600 | Major — reduced to primary planes | Yes, as a filler prop |
hard surface, beveled edges |
52,400 | Moderate — panel seams and lid rim now real geometry | After decimation |
chunky proportions, flat bottom |
38,900 | Moderate — thicker walls, stable footprint | After decimation |
battle damaged, dented lid |
44,100 | Moderate — genuine dents in the silhouette | After decimation |
octane render, 8k, cinematic lighting |
41,500 | None — plus baked highlights in the albedo | No, and now the texture needs fixing |
Two rows in that table cost you nothing and gave you nothing. That ratio is roughly what you should expect from prompts written out of habit rather than intent.
Iterate in preview, not in refine
Text-to-3D splits into a cheap geometry preview and a more expensive textured refine. Every geometry experiment belongs in preview. Generate four or five preview variants at 5 credits each, judge them purely on silhouette and proportion with textures off, and only refine the one you are keeping. Refining three candidates to compare them is a straightforward way to spend four times what you needed to.
Judge previews in flat grey shading, rotated. A prompt that looks good in a thumbnail often has a back you would never ship, and untextured grey is the honest view.
Do and don’t
- Do change one phrase at a time when testing. Changing three and liking the result teaches you nothing.
- Do keep a personal phrase book of what worked, with a thumbnail and the triangle count next to each entry.
- Do describe what the object is for. “For wall mounting”, “sits on a desk”, “held in one hand” all shift proportion in useful ways.
- Don’t stack more than two style words. “Low poly stylized cartoon hand surface realistic” cancels itself out into mush.
- Don’t use negatives for things you never mentioned. Asking for “no text, no logo” on an object that has neither adds noise.
- Don’t prompt for material names expecting geometry. “Woven wicker” gets you a wicker texture on a smooth basket, not woven strands.
- Don’t keep regenerating an identical prompt hoping for a better roll. After three tries, the prompt is the problem.
Build your own phrase book this week
Pick one object you understand well — a crate, a helmet, a lamp. Write a bare five-word prompt and generate it. Then run eight more previews, each adding exactly one phrase from the table above, and screenshot every result in flat grey at the same camera angle. In an afternoon and under a hundred credits in a studio like MeshyFlix, you will have a personal reference sheet showing what each phrase does in the shapes you actually make, which is worth considerably more than any generic list. Keep it next to you, add a row every time something surprises you, and stop guessing.