All news

Writing prompts that produce better 3D models

September 23, 2026 · 5 min read

A good 3D prompt describes one object: its shape, its materials and its style. Here is a structure that works, what to leave out, and how to iterate without wasting credits.

A text prompt on a desk dissolving into triangles that form a knight's helmet

Text-to-3D models read your prompt differently from image generators. An image model has to invent a camera, a background, a light setup and a composition. A 3D generator has none of that: its output is a single object that you will later light, frame and place yourself. The words that help an image prompt — "cinematic", "golden hour", "shallow depth of field" — mean nothing to a mesh, and at worst they add noise that pulls the shape in odd directions.

This guide covers a prompt structure that works well for single objects, the words that are better left out, how the negative prompt fits in, and a way to iterate that keeps your credit spend predictable.

Start from the object, not the scene

The single most useful habit is to name one object in the first few words. "A medieval knight's helmet" gives the generator a clear subject. "An epic scene of a knight guarding a castle at dawn" gives it a knight, a castle, a sky and a mood, and you will get one mesh that tries to be all of them at once.

If you really need several things together — a lamp on a desk, a sword in a stand — describe them as one object ("a desk lamp mounted on a small wooden base") and accept that they will come out as a single connected mesh. For separate objects, generate them separately and assemble the scene in your engine or modelling tool.

A structure that works

Most good prompts follow the same four-part order. You do not need every part every time, but when a result disappoints, the missing part is usually the reason.

  1. Subject. What the object is, in plain words: "a brass diving helmet", "a low stool", "a cartoon fox".
  2. Shape and proportions. The silhouette is what the generator most often gets wrong, so describe it: "round dome, three circular portholes, heavy collar", "short legs, wide round seat", "oversized ears, bushy tail curled to one side".
  3. Materials and surface. "Worn brass with green patina and riveted seams", "oak with visible grain", "soft orange fur with a white chest". Materials drive both the texture and, with PBR texturing, the roughness and metalness maps.
  4. Style. "Realistic", "stylised", "low-poly game asset", "toy-like". One or two style words are enough; stacking five adjectives rarely helps.

Put together:

A brass diving helmet, round dome with three circular portholes and a heavy bolted collar, worn brass with green patina, realistic.

That prompt is short, but every phrase carries shape or material information the generator can use.

Describe what you would see from all sides

A 3D model has a back. If the back matters — a backpack on a character, a handle on a mug — mention it. Generators fill in hidden sides with plausible guesses, and a guess that contradicts what you wanted is the most common reason for a second attempt.

Use concrete, countable details

"Four legs", "two straps", "a single handle on the left" are easier to follow than "some legs" or "decorative details". Numbers and positions give the generator anchors.

What to leave out

Leave out Why Instead
Backgrounds and environments You get one mesh; extra nouns become extra geometry Generate the environment pieces separately
Camera words (close-up, wide angle, from above) There is no camera in a 3D model Frame the model in the viewer or your engine
Lighting words (dramatic, sunset, neon glow) Lighting belongs to the scene, not the asset Choose lighting when you render
Long lists of adjectives They compete with each other Keep one or two style words
Text, logos, lettering Small text rarely survives as geometry or texture Add decals or text later in a 2D tool

The negative prompt

The negative prompt is for things you keep getting by accident. Typical entries are "base, pedestal, text, extra limbs, floor". Keep it short and specific: it works best as a list of a few unwanted features, not as a second description. In MeshForge the negative prompt accepts up to 255 characters and is available for text prompts.

Choosing settings alongside the prompt

The prompt is only half of the request. A few studio settings change the result as much as the words do:

  • Model. The standard model aims for detailed geometry; the low-poly options aim for game-ready meshes with far fewer faces.
  • Texture. None gives you an untextured mesh, which is the cheapest way to check a shape. Standard gives a colour texture; PBR adds roughness, metalness and normal maps for engines and renderers.
  • Polycount. A target triangle count. Leave it adaptive unless you know your budget — see our guide on polycount.
  • Seed. The same seed and the same input give the same mesh, which is useful when you want to change only one thing.

Every one of these options shows its credit price before you start, so you can see what a change costs before you commit to it.

Iterate one change at a time

When a result is close but not right, resist rewriting the whole prompt. Change one part — the shape description, or the material, or the style word — and keep the seed the same. That way you learn what each part of the prompt actually does.

A practical loop:

  1. Generate an untextured preview to check the silhouette.
  2. Adjust the shape words until the silhouette is right.
  3. Switch on texturing (Standard or PBR) once the shape works.
  4. Use Remix on a result to reopen it in the studio with the same prompt and settings, ready to tweak.

Worked examples

Too vague: "a cool robot". The generator has to guess everything: size, proportions, materials, style.

Better: "a small round cleaning robot, dome-shaped body on three wheels, one camera eye, white plastic with orange trim, stylised". Subject, shape, materials and style are all present.

Too crowded: "a wizard casting a spell in a dark forest with glowing runes and a raven on his shoulder, dramatic lighting". Background, lighting and a second creature are all fighting for the same mesh.

Better: "a wizard in a long hooded robe holding a gnarled staff, raven perched on his left shoulder, dark blue robe with silver embroidery, stylised". The raven stays because it is part of the silhouette; the forest and lighting go.

Sources & further reading