ColorYourModel Fuse / Recommend Stage Benchmark
Release-mode stage timings for the one-click fuse pipeline (fuse_segmentation)
and seed recommendation (recommend_seeds), measured by the
bench_fuse_and_recommend_stages test in src-tauri/src/segment/mod.rs.
What it measures
The bench mirrors the fuse_segmentation command stage by stage —
planar (15/M30) → multiview (12 views) → dihedral fold → edge vote with the
SeedPanel's real cut=2/min=0 — plus recommend_seeds at panel defaults
(count 12, curvature 1.0, concavity 1.0). A fresh load per CYM_BENCH_DIHEDRAL
setting matches the command path. It reports per-stage timings, region counts,
and max/tiny-region shares.
How to run
# Release mode; the test is #[ignore] because it reads large STL files.
CYM_BENCH_STL=samples/_src/YourModel.stl CYM_BENCH_DIHEDRAL=15,2 \
cargo test --release --lib bench_fuse_and_recommend_stages -- --ignored --nocapture
CYM_BENCH_STLselects the model (default:samples/_src/Sanji+Diorama+Detailed_U1.stl).CYM_BENCH_DIHEDRALis a comma list of fold-angle settings (default15,2).
Models are NOT shipped with the repository. Drop your own STLs under
samples/_src/ (gitignored) or point CYM_BENCH_STL anywhere on disk.
Results (2026-09-21 session, 17 STLs, release build)
Per-model details were recorded for three representative models:
| Model | Faces | Fold angle | planar | multiview | dihedral | vote | Total | Regions |
|---|---|---|---|---|---|---|---|---|
| godzilla | 1.5M | 2° | 0.4s | 9.4s | 2.2s | 0.5s | ~12.7s | 70 |
| luffy | 500K | 2° | — | — | — | — | 8.0s | 87 (max region share 8.0%) |
| house2 | 4.7M | — | — | — | — | timeout | >30min | — |
Scalability finding
On house2 (4.7M faces) the vote stage ran over the 30-minute bench timeout
while iterating 4.73M dihedral sets: MAX_AUTO_REGIONS (postprocess.rs,
4096) is not wired into the fuse path, so nothing caps the work on huge
meshes. Wiring the cap into fuse_region_sets is recorded as backlog.