40-run teardown of Trellis.2: stacking Cumesh and Meshlib simplifiers is the winning combo

Tamerygo · reddit · 2026-09-06

After Trellis.2 went native in core, the author ran 40 controlled experiments on the 3D generation pipeline and found the two simplify stages do fundamentally different jobs, and order matters:

Cumesh is unreliable but topology-friendly: on a 300k-triangle budget, the same graph produced 75k, 1.88M, and 4.79M faces across runs; a 2.4M-face mesh then cost 79s per texture step.
Meshlib hits targets exactly but shreds topology: always exactly 250,000 triangles, with 90× more holes.
The key combo: the hole-filling node sits between the two stages, so Cumesh → fill → Meshlib yields 1,261 holes, 8s of fill time, and exactly 250k triangles — now the default.
Reconstruction grid: lower resolution isn't just faster, it leaves far fewer holes for the fill stage to patch one at a time.
On 16 GB cards: 64³ detail grid is 8× the decode of 32³, pages into system memory and appears to hang for minutes; 32³ is the right call.
A disguised failure mode: concept images with objects touching the frame edge give the background remover no silhouette, producing flat cards instead of meshes — auto-padding margins fixed it.

The author also built a local orchestration tool that sends generated models to Houdini and Unreal.

Original post →

40-run teardown of Trellis.2: stacking Cumesh and Meshlib simplifiers is the winning combo

More from Multimodal