A prompt for 3D generation is describing an object, not a picture of a scene. The habits that make a good illustration prompt actively hurt here.
Include
- The object, named plainly. "Bar stool", "longsword", "ceramic teapot". Nouns do the heavy lifting.
- Material. "Brushed steel", "oak", "glazed ceramic". Materials change both the shape read and the surface.
- Defining features. "Three legs", "curved back", "hexagonal handle" — anything that distinguishes this object from the generic version.
- Style, if it matters. "Low poly", "stylised", "realistic". One word is enough.
- A plain background. Explicitly, if the generator writes an image first.
Leave out
- Lighting. "Dramatic rim light", "golden hour" — these change the picture and not the object, and strong lighting actively harms reconstruction by baking shadows into shape.
- Camera and lens. "85mm", "shallow depth of field" — blur removes information the reconstructor needs.
- Mood and quality words. "Beautiful", "masterpiece", "8k", "trending". They cost tokens and change nothing about the geometry.
- Scenes. "A sword resting on a table in a tavern" produces a mesh containing a sword, a table fragment and some tavern.
One object per prompt
This is the rule that catches people most often. A generator asked for two things produces one mesh containing two things, fused where they touched. If you need a set, generate the pieces separately.
Iterating
Change one thing at a time. Prompt behaviour is not linear, and a rewritten prompt tells you nothing about which word mattered. If a result is close, adjust the single term that describes what is wrong.
If you find yourself writing longer and longer prompts to pin down a shape, that is the signal to switch to image to 3D — a picture specifies a silhouette in a way a sentence never will.