Guides
How to write prompts for AI 3D model generation
How to write text prompts for AI 3D model generators: a simple structure, words that help and words that hurt, and how to refine a result.
3 min read
Quick answer
How do you write a good prompt for an AI 3D generator?
Describe one object, not a scene. Name what it is, its main shape, its materials and its style, in that order. Leave out camera angles, lighting and backgrounds, because the generator builds a model rather than a picture. Keep the object in a neutral pose, and change one word at a time when refining.
A 3D prompt is not an image prompt
If you have used an image generator, you have probably learned to write prompts like a photographer: "cinematic lighting, shallow depth of field, golden hour, 85mm". For a 3D model, most of that is noise or worse. There is no camera. The lighting will come from whatever scene you put the model in. A blurred background has nowhere to go.
A 3D prompt describes a thing, so write it like a description for a prop-maker.
A structure that works
Put the parts in this order, and stop when you have said enough:
- What it is. "A wooden treasure chest."
- Shape and proportions. "Wide and low, with a curved lid."
- Materials. "Dark oak planks, iron bands and corner brackets, a brass lock."
- Condition. "Weathered, scuffed edges."
- Style. "Stylised, hand-painted look" or "realistic".
Together: a wooden treasure chest, wide and low with a curved lid, dark oak planks with iron bands and a brass lock, weathered edges, stylised hand-painted look.
Words that help
- Materials by name. "Brushed steel", "glazed ceramic" and "worn leather" give the texture stage something concrete.
- Proportion words. "Squat", "tall and narrow", "chunky", "slender".
- A style anchor. "Low-poly style", "stylised" and "realistic" steer the whole result. Keep the same anchor across a project so assets match.
- "Game asset" or "single object". These tell many generators to produce a standalone object.
Words that hurt
- Scenes. "A tavern with tables and a fireplace" asks for many objects. Generators do far better with one, so generate the table, the chair and the fireplace separately.
- Camera and lighting terms. "Close-up", "dramatic lighting" and "bokeh" can end up baked into the textures as shadows and highlights.
- Action poses. "A knight swinging a sword" gives you a static mesh frozen mid-swing that is hard to rig. For anything you will animate, ask for a neutral pose, such as standing with arms out.
- Text and logos. Lettering rarely comes out readable on 3D models. Add it afterwards as a decal or texture.
How to refine
- Generate several, then pick on silhouette. Shape is hard to fix later. Colour is easy.
- Change one thing at a time. If you change four words and the result improves, you will not know which word did it.
- Switch input when words run out. If you keep writing "but more like…", draw it instead. Sketch or image input carries shape far better than adjectives.
Example prompts to adapt
- A sci-fi supply crate, rectangular with chamfered corners, painted metal with yellow hazard stripes, scratched paint, game asset.
- A ceramic table lamp, round base, conical linen shade, matte white glaze, minimalist product design, realistic.
- A potion bottle, round body and long neck, green glass with a cork stopper and leather cord, stylised.
These describe what to ask for, not what any tool will return. Results vary with every generator and every run.
Vi3W's text to 3D follows this structure well, and because every result is retopologised automatically, you can judge candidates on shape and material rather than on how messy their wireframe is.
Last reviewed