H01
Uniform quadric collapse
Clean baseline. Fastest safe default, but with no explicit boundary guard.
FORGEFX / 3D ASSET OPTIMIZATION STUDY
Grader Pedals: Ten Ways Under 50K
Ten real geometry-reduction hypotheses were executed against the supplied FBX. Nine preserve the visible silhouette at this review angle; one destroys mechanical fidelity. Meshy and Unity Asset Transformer were also probed directly.
Source geometry225,130triangles / 116,072 vertices
Target result48,000triangles; 78.68% reduction
Best image SSIM0.999509material / UV / normal guard
Source materials27preserved by all Blender finalists
RECOMMENDATION
Use H02 as the immediate 48K screening winner.
It reached 48,000 triangles, retained all 27 material slots after FBX re-import, produced the best measured render similarity, and showed no visible silhouette loss. For production, improve it with component-aware cleanup: the source packs 489 disconnected CAD shells into one mesh, and blind welding can damage intended hard edges and separations.
Baseline And Overall Comparison
Original source. One mesh object, 27 materials, 225,130 triangles. The uploaded FBX was not modified.
Baseline followed by H01–H10, left-to-right and top-to-bottom. H08 voxel remesh is the obvious failure; the other nine are visually near-identical at this view.
Why The Source Is Expensive
Verified wireframe proof from Blender 5.1. The mesh is fully triangulated CAD tessellation with concentrated density around small mechanical details.
SOURCE TOPOLOGY
One object contains 489 disconnected shells
That structure explains why broad decimation can look good from one angle while still threatening holes, fasteners, tread, pivots, and shell boundaries.
Only 6 of 27 material slots contain triangles. A production pass should remove the 21 unused slots, split or classify meaningful shells, strip hidden/internal geometry, then allocate triangle budgets by part importance. Do not blindly weld duplicate positions; some preserve hard normals or intentional CAD seams.
The Ten Executed Hypotheses
H02 · WINNER
Material-boundary guard
Best balance of measurable fidelity and production-safe intent. Recommended.
H03
UV-seam guard
Matched H02 visually and numerically at this target. Useful when texture seams are the priority.
H04
Sharp-normal guard
Matched H02/H03 at this view. Sensible hard-surface safeguard.
H05
Planar dissolve at 2°
Conservative coplanar cleanup before collapse. Passed, but added no visible advantage.
H06
Planar dissolve at 8°
Still visually acceptable, but the lowest similarity among the non-destructive finalists.
H07 · RUNNER-UP
Micro-weld, then collapse
Fewest vertices among the 48K finalists and 0.999501 SSIM. Validate shell intent before production use.
H08 · REJECT
Voxel topology reset
Massive over-reduction. Merged and deformed mechanical parts; unsuitable.
H09
Unsubdivide, then collapse
Passed, but weaker than direct collapse and unnecessary for this triangulated source.
H10
X-symmetry-aware collapse
No measurable benefit over H01. The source is not symmetric enough for this to matter.
Verified Result Table
| Hypothesis | Triangles | Vertices | Reduction | FBX size | Materials after re-import | SSIM | Decision |
|---|---|---|---|---|---|---|---|
| H01 Uniform | 47,999 | 26,940 | 78.68% | 3.11 MB | 27 | 0.999491 | PASS |
| H02 Material guard | 48,000 | 27,021 | 78.68% | 3.10 MB | 27 | 0.999509 | RECOMMEND |
| H03 UV guard | 48,000 | 27,021 | 78.68% | 3.10 MB | 27 | 0.999509 | PASS |
| H04 Normal guard | 48,000 | 27,021 | 78.68% | 3.10 MB | 27 | 0.999509 | PASS |
| H05 Planar 2° | 47,992 | 26,940 | 78.68% | 3.39 MB | 27 | 0.999412 | PASS |
| H06 Planar 8° | 47,984 | 26,960 | 78.69% | 3.44 MB | 27 | 0.998584 | PASS |
| H07 Micro-weld | 48,000 | 23,963 | 78.68% | 3.09 MB | 27 | 0.999501 | RUNNER-UP |
| H08 Voxel | 3,292 | 1,650 | 98.54% | 0.18 MB | 27 | 0.953037 | REJECT |
| H09 Unsubdivide | 47,998 | 26,699 | 78.68% | 3.12 MB | 27 | 0.999105 | PASS |
| H10 Symmetry X | 47,999 | 26,940 | 78.68% | 3.11 MB | 27 | 0.999491 | PASS |
Requested Toolchain Probes
MESHY REMESH API · EXECUTED
47,930 triangles, but visually unacceptable
Meshy completed a real 48K triangle remesh. It collapsed 27 material slots to one, normalized the longest dimension to 2 m, introduced about 0.5% non-uniform proportion drift, and visibly melted holes, tread, fasteners, plates, and linkages. Recognizable silhouette; failed mechanical fidelity.
UNITY ASSET TRANSFORMER / PIXYZ · ATTEMPTED
The package test hit disk capacity before resolution
A clean Unity 6.1 project was created successfully and the supplied FBX was staged. The live floating license reported Pro access. Adding [email protected] reached Package Manager, then stopped with ENOSPC while only 1.1 GB remained. Asset Transformer therefore produced no optimization result in this pass.
“Draft Pixi” was treated as Pixyz, the former name of Unity Asset Transformer. The current package could not be judged until adequate disk space is available.
Evidence standard. Every Blender FBX was exported, re-imported into Blender 5.1, and recounted. Render similarity is whole-frame SSIM against the source from the same orthographic review camera. This is strong screening evidence, not a substitute for close-up UV, collision, pivot, normal-map, and engine tests. The source checksum is
c3c9cb3f3248a11eb9894a83c8a71c008207661b3d928c3f0f5b9419b69128c7.Source preserved unchanged. Review generated July 31, 2026. Preview is link-accessible and potentially indexable; it contains screenshots and metrics, not the FBX files.