Short answer: generation is a drafting tool, not a replacement for modelling. It wins decisively where you need many plausible objects quickly, and loses where a shape has to be correct rather than convincing. The interesting question is not which to use, but where the handover sits.
| AI generation | Modelling by hand | |
|---|---|---|
| Time to first result | Seconds to minutes | Hours to days |
| Cost per asset | Low, and roughly flat | High, and scales with the artist |
| Exact dimensions | Not guaranteed | Guaranteed |
| Topology | A reconstruction — needs rebuilding | Built deliberately |
| Consistency across a set | Hard | Straightforward |
| Revisions | Regenerate, take what you get | Edit precisely |
| Best at | Volume, drafts, background, exploration | Hero assets, mechanical parts, anything specified |
What generation is genuinely good at
Volume. Forty crates, thirty rocks, a street of buildings that share a language. The economics invert here: a human modelling forty variations is a week, and the fortieth is worse than the first because attention runs out.
The first draft. A generated mesh you then retopologise, unwrap and re-texture is often faster than starting from a cube, and the silhouette decision — usually the hardest part — is already made.
Exploration. Deciding what a thing should look like is cheaper in a generator than in a modelling package, because you are not invested in the last twenty minutes of work.
What it is genuinely bad at
Anything with a specification. A bracket that must accept an M6 bolt, a shelf that must be 400 mm deep, a part that must mate with another part. Reconstruction infers shape from appearance and appearance does not carry dimensions.
Interiors and hidden structure. A generated mug is a solid shape that looks like a mug. Opening a drawer that was never modelled as a drawer reveals a hole. Segmentation divides the surface; it does not invent what was behind it.
Consistency. Two generations from two similar prompts are two unrelated meshes. A modular kit whose pieces must snap together on a grid is a hand-modelling job, or a generation job followed by a lot of hand correction.
Topology, always. Reconstruction produces dense, irregular triangles with no edge flow. That is not a defect to complain about, it is what the technique produces — see why generated meshes need rebuilding. Anything that will deform, subdivide or be textured properly needs retopology first.
The pipeline most people actually end up with
Generation for the draft and the volume; hand work for the parts that carry the scene. Concretely:
- Generate the silhouette and the base shape.
- Retopologise to a polygon budget you chose rather than one the algorithm produced.
- Unwrap, because a new mesh has no usable UV layout.
- Hand-correct the parts a viewer will look at closely — and only those.
The judgement call is which assets deserve the hand-correction pass. On most projects it is a small minority of them, which is precisely why the volume case matters.
What this does not settle
Whether generated output is appropriate for your project is a licensing and disclosure question as much as a technical one, and it varies by client, platform and jurisdiction. The technical comparison above says nothing about it.