A level of detail is a simplified version of a mesh, swapped in as the object gets further away. The technique is old and the decisions around it are still where most of the value is.
How many levels
Three or four is typical for a prop. More than that and each level is doing too little to justify the memory it costs.
A common progression is full detail, then roughly half, then a quarter, then a silhouette-only version — with the final level often being an impostor or nothing at all.
Where to switch
Screen coverage, not world distance. A large object at fifty metres occupies more pixels than a small one at ten, and it is the pixel count that determines whether detail is visible. Engines express this as a screen-size percentage.
Making the pop invisible
- Match the silhouette. Reduction that changes the outline is what the eye catches. Interior detail can go; the profile cannot.
- Switch when the object is small. Most popping complaints are switch thresholds set too aggressively.
- Preserve UV layout. A reduced mesh that re-unwraps will shift its texture visibly at the switch.
- Consider dithered transitions. Cross-fading between levels over a few frames hides what an instant swap reveals.
When LODs are the wrong tool
- Heavily instanced small props. The draw call cost dominates the vertex cost; batching helps and LODs barely do.
- Objects always seen close. A weapon in first person has one distance.
- Scenes bound by overdraw or texture memory. LODs reduce vertices. If vertices are not your bottleneck, they buy nothing.
Generated meshes
Reconstruction density is arbitrary, so the "full detail" level is usually already wrong. Retopologise to your intended LOD0 first, then generate the chain from that — not from the raw output.