Image to 3D reconstructs geometry from a single picture. It is the most predictable way to get a specific object, because you are not asking a model to invent the object first — only to work out its shape.
What makes a good input
The reconstructor reads a silhouette and infers depth from shading. Anything that confuses either one costs you geometry.
- Isolate the subject. A plain background is worth more than a high resolution one. Background removal usually improves the result outright.
- Show the whole object. A cropped edge becomes a cut-off mesh, not an inferred continuation.
- Even, diffuse light. Hard shadows read as shape. A dark crease under a chair seat can reconstruct as an actual crease in the geometry.
- Avoid extreme perspective. A wide-angle photo taken close up distorts proportions, and the mesh inherits the distortion.
- Three-quarter views beat straight-on. A front-on photo shows almost no depth information; turning the object thirty degrees gives the reconstructor two visible faces to work from.
What the back looks like
Whatever the model thinks is plausible. For a symmetrical object — a vase, a wheel, a bottle — that guess is usually good. For anything with a distinct rear, it is invention, and it will look like invention.
If the back matters, use multi-view to 3D and give it real views.
Transparent and reflective objects
Both are hard for the same reason: what the camera recorded is not the surface. Glass shows what is behind it and chrome shows the room. The reconstructor has no way to separate the object from its reflection, so it tends to produce lumpy, over-detailed geometry with the reflection baked into the texture. Matte-paint the object and re-photograph it if you can; otherwise expect to clean up.
After the mesh
A reconstructed mesh is dense and its topology is machine-made — fine for a render, awkward for animation or further modelling. Retopology rebuilds it with cleaner geometry, and UV unwrapping gives it a layout you can paint on.