How to write text-to-3D prompts that produce clean models
Text-to-3D rewards a different kind of prompt than image generation. Here is the structure that works and the habits that waste credits.
By the threeD editorial team. Published . 7 min read.
If you have used image generators, your instinct will be to write long, atmospheric prompts full of lighting and camera words. Text-to-3D works differently. There is no camera and no scene: the generator is building one object that will be lit later by whatever engine or viewer you put it in. Words about mood and lighting mostly get ignored, or worse, get baked into the texture as painted-on shadows.
What matters is what the object is, what it is made of, and what state it is in.
The four-part structure
A reliable order for a text-to-3D prompt is:
- Subject: the object itself, named as specifically as you can. “A wooden treasure chest” beats “a box”.
- Materials: what each main part is made of. “Oak planks, iron bands and a brass lock.”
- Condition: new, worn, rusted, chipped, mossy, polished. This mostly affects texture, but heavy damage can change the shape too.
- Style: realistic, stylised, low-poly, hand-painted. Put it last, because it is the part you are most likely to change between attempts.
Put together:
A wooden treasure chest, oak planks with iron bands and a brass lock, scuffed and slightly mossy, stylised hand-painted look
You can see this structure applied to sample models on the text to 3D page, where each part of the prompt is highlighted.
One object per prompt
The most common reason for a messy result is asking for a scene. “A desk with a lamp, a laptop and a coffee mug” usually becomes a single fused mesh where the mug melts into the desk. Generate each object separately and arrange them in your engine or editor. It costs a little more, but each piece is clean, reusable and can be optimised on its own.
Small attached parts are fine. “A lamp with a pull chain” is one object. “A lamp next to a book” is two.
Use the negative prompt for repeat offenders
Generators have habits learned from their training data. Objects often appear on pedestals or bases; characters sometimes get a ground plane; signs and labels get garbled text. If you see the same unwanted feature twice, add it to the negative prompt instead of rewording the main prompt. Common entries:
no base, no pedestal, no groundfor propsno text, no lettersfor anything with a surface that might carry a labelno baked shadows, no lightingwhen the texture looks darker on one side
Characters: ask for the pose you need later
If a character will be rigged, the pose at generation time decides how well auto-rigging works. Arms touching the body or legs pressed together cause skin weights to bleed between parts. Ask for it explicitly:
A cartoon fox adventurer in a leather jacket, full body, T-pose, arms straight out, legs apart, facing forward
“Full body” matters too. Without it, some generations crop at the waist. The rigging page explains why the pose matters in more detail.
Words that help and words that do not
| Usually helps | Usually ignored or harmful |
|---|---|
| Materials: “ceramic”, “brushed aluminium”, “knitted wool” | Lighting: “golden hour”, “dramatic lighting”, “studio lit” |
| Shape words: “chunky”, “tall and narrow”, “rounded edges” | Camera: “close-up”, “wide angle”, “85mm” |
| Era or design language: “art deco”, “1970s”, “Scandinavian” | Quality boosters: “8K”, “masterpiece”, “trending” |
| Counts: “four legs”, “two handles” | Other objects in the scene |
Iterate on the preview, not the refined model
Previews are cheap and show you the shape. If none of the four previews has the right silhouette, change the prompt and run previews again. Only refine once the shape is right; texture problems can be fixed afterwards with AI texturing without regenerating the mesh.
A practical loop:
- Run previews with a short prompt: subject and main material only.
- If the shape is wrong, add shape words or counts. If parts are missing, name them.
- Once one preview looks right, add condition and style and refine it.
- Fix texture details with a texturing pass instead of starting over.
A prompt checklist
- Is there exactly one object?
- Is the subject named specifically?
- Does every main part have a material?
- For characters: full body, T-pose or A-pose, facing forward?
- Have you removed lighting, camera and quality-booster words?
- Is anything you keep seeing and do not want in the negative prompt?
If you are starting from a photo or drawing instead of words, read the companion guide on choosing a reference image.