How to write text-to-3D prompts for game assets
Updated
A text-to-3D prompt has a different job from an image prompt. You're describing an object that has to hold up from every angle, so mood, camera, and scenery matter far less than form, proportion, and material.
These are general starting principles, not findings from a benchmark. Generators differ, so treat each one as something to try on your own assets.
Describe one object
Ask for a single thing rather than a scene. If you need a table with a lantern on it, you probably want two separate assets anyway, so you can move, reuse, and optimize them independently.
Follow a simple order
A consistent structure makes prompts easier to compare and adjust:
- Subject: what it is, such as a wooden treasure chest.
- Form: shape and proportions, such as squat with a rounded lid.
- Materials and color: such as worn oak with golden metal trims.
- Style: such as stylized, low-poly, or realistic.
- Use: such as a prop for a colorful fantasy game.
Put together, that might read:
A stylized low-poly wooden treasure chest with golden metal trims, suitable for a colorful fantasy game.
Leave out what a model can't use
Camera angles, lens choices, dramatic lighting, and backgrounds belong to image generation. At best they do nothing here. At worst the lighting gets painted into the texture, which then looks wrong in your game. It's often worth asking for even, neutral lighting instead.
Plan for the back and the underside
If a detail matters, say where it is. "A backpack with a rolled blanket strapped to the back" gives the generator more to work with than "a detailed backpack". Vague detail tends to produce vague surfaces in the places you look at last.
Pose characters for rigging
If a character will be animated, ask for a neutral pose, such as a T-pose or A-pose, with arms and legs clear of the body. A dramatic pose may look better in the preview, but limbs that touch or merge into the torso are harder to rig.
Change one thing at a time
When a result is close, change a single phrase and generate again. If you change subject, style, and material together, you won't learn which change helped. Keep a short log of prompts and results so you can repeat what worked.
When to use an image instead
If you already know how something should look, a reference image may get you closer than words. Clear inputs tend to help: one object, a plain background, an uncluttered silhouette, and an angle that shows its three-dimensional form. Make sure you have the right to use any image you upload.
Check the result
A good prompt gets you a good draft. Before you build on it, run it through our inspection checklist.