Text-to-3D and Image-to-3D Prompting

Text prompts for 3D

Describe the object, not a scene:

A stylised ceramic panda figurine sitting upright, holding a small laptop,
smooth rounded shapes, matte glazed finish, black and white with a soft green laptop

• Name the object, its pose or configuration, materials and style (realistic, low-poly, cartoon, sculpted).
• Avoid backgrounds, lighting and camera terms: they do not apply to a standalone model.
• Mention what matters for the use: "thick, sturdy parts" for 3D printing, "low-poly game asset" for real-time use.

Preparing images for image-to-3D

• One centred object, plain background, even lighting, minimal shadows.
• A three-quarter view usually works better than a straight-on view.
• Multiple views (front, side, back) improve accuracy where supported.
• Remove the background first if needed.
• Use images you have the rights to.

Iterating

1. Generate several candidates.
2. Inspect each from all angles: the back and underside are often weakest.
3. Refine the prompt or image and regenerate, or pick the best base to clean up.
4. Generate or refine textures separately once the shape is right.

Settings to know

• Polycount / target faces: lower for games and the web, higher for detailed prints.
• Topology options (triangles vs quads) where available.
• Symmetry for characters and many objects.
• Texture resolution and PBR maps (base colour, roughness, metallic, normal).

Exercise

Generate the same object three ways: text-to-3D, image-to-3D from a photo you took, and image-to-3D from a generated concept image. Compare shape accuracy, back-side quality and textures, and note which inputs worked best.

Check yourself

• Why describe an object rather than a scene in a 3D prompt?
• What makes a good reference image?
• Which part of generated models is often weakest?