Asset estimates go wrong because "a 3D model" describes work spanning two orders of magnitude. Estimating means classifying first.
What drives the number
- Role. A background prop and a hero character are different professions, not different sizes.
- Whether it deforms. Rigging and skinning roughly change the shape of the job.
- Reference availability. A clear brief with reference is far faster than an ambiguous one.
- Iteration expectations. Rounds of feedback often exceed the initial build.
- Target constraints. A mobile budget requires more optimisation work than a desktop one.
- Whether it must match something exactly. Approximation is fast; specification is not.
A rough tiering, in effort rather than currency
- Simple static prop, clear reference, no deformation: a few hours.
- Detailed prop with baked detail and authored materials: a day or so.
- Hero prop, art-directed, multiple iterations: several days.
- Character, modelled, textured, rigged: weeks rather than days.
- Modular kit, designed to tile: days, and it repays itself across the environment.
These are ranges for an experienced artist with clear direction, and the range widens considerably without either.
What people forget to include
- Reference gathering and briefing
- Iteration rounds
- LODs and collision
- Engine integration and testing
- Fixing what integration reveals
The modelling is frequently less than half the total.
With generation
The build stage compresses substantially; the finishing stages do not. So the saving is real and it is a saving on part of the job — estimating a generated asset as "free" is the error that produces disappointing schedules.
A reasonable planning assumption for suitable assets is a meaningful reduction on the build, with cleanup and integration unchanged.