Automatic Skinning Benchmark: All Methods Compared | Jessy Leite
This is the cross-method view of the skinning benchmark: four automatic methods run on the same 202 characters, each scored against the weights the artists painted by hand on the same meshes. Every method runs the same protocol, on the same characters, with the same ground truth, on the same MacBook Air M4. Each method also has its own page in this series with the full per-character tables (Adaptive Auto Skinning<br>, Voxel Heat Diffuse<br>, Automatic Weights<br>); this article is the synthesis.
The methods: Blender’s built-in Automatic Weights (bone heat), the Voxel Heat Diffuse Skinning addon (Mesh Online) and its older Surface Heat Diffuse solver, and the volumetric binder I am building, Adaptive Auto Skinning<br>. How the binder’s refinement works<br>is written up separately. Maya’s Geodesic Voxel Binding needs a Maya license to score and is not in yet.
Results
Every percentage is how close the method’s weights are to the artist’s, so higher is better: pp is the dominant-bone match (top-1), tv the full weight-vector similarity, and btv the per-bone average (a finger counts as much as the spine). Speed is the bind time. Full definitions are at the end of the page
Mixamo (102 characters)
Method<br>Success<br>pp (%)<br>tv (%)<br>btv (%)<br>Speed (s)
Adaptive Auto Skinning<br>102 / 102<br>80.9<br>79.5<br>76.0<br>0.57
Voxel Heat Diffuse (res 128)<br>102 / 102<br>75.2<br>71.4<br>63.0<br>1.87
Voxel Heat Diffuse (res 368)<br>102 / 102<br>76.0<br>74.9<br>72.4<br>139.75
Surface Heat Diffuse<br>102 / 102<br>75.0<br>74.7<br>74.1<br>0.34
Automatic Weights (bone heat)<br>58 / 102<br>38.8<br>39.6<br>41.9<br>0.14
RigNet (100 characters)
Method<br>Success<br>pp (%)<br>tv (%)<br>btv (%)<br>Speed (s)
Adaptive Auto Skinning<br>100 / 100<br>74.3<br>71.5<br>64.9<br>0.55
Voxel Heat Diffuse (res 128)<br>100 / 100<br>70.7<br>68.9<br>61.6<br>3.36
Voxel Heat Diffuse (res 256)<br>99 / 100<br>69.7<br>69.0<br>61.0<br>59.82
Voxel Heat Diffuse (res 368, partial)<br>28 / 28<br>69.5<br>69.3<br>59.7<br>275.14
Surface Heat Diffuse<br>100 / 100<br>69.4<br>68.8<br>60.6<br>0.09
Automatic Weights (bone heat)<br>85 / 100<br>58.2<br>58.0<br>51.8<br>0.01
Voxel Heat Diffuse was run at 128 and 368 on Mixamo, and 128, 256, and a partial 368 on RigNet (only 28 of the 100 characters finished at 368 before the run was stopped). Higher resolution costs far more time for no real quality gain.
Top-left is better. Each point is the mean over Mixamo and RigNet. The Voxel Heat Diffuse line is its 128 to 368 resolution sweep: bind time explodes while quality stays flat. (VHD 256 is RigNet only and VHD 368 on RigNet is partial, which is why 256 reads a touch low.)
Weight quality against the artist, three metrics, Mixamo.
What stands out
Four things, each with its caveat:
Adaptive Auto Skinning tops every quality column on both datasets and binds in about half a second. These are its first numbers, on one benchmark, and it has not been stress-tested across the long tail of production geometry the way the others have.
Voxel Heat Diffuse is the reliable workhorse. It skinned every character with no failures at solid quality, and the cost is CPU time. Raising its voxel resolution from 128 to 368 adds a few points of weight-vector similarity for a 75 times longer bind (1.9 s to 140 s on Mixamo), so the default resolution is usually the right one. The Voxel Heat Diffuse benchmark page<br>has the full resolution sweep.
Surface Heat Diffuse is fast and never failed. The older Mesh Online solver bound all 202 characters in under half a second each, matched Voxel Heat Diffuse at resolution 128 on Mixamo dominant match (75.0% against 75.2%), and kept small-bone weights better (btv 74.1% against 63.0%). On the non-humanoid RigNet rigs it slips a few points under Voxel Heat. It is the same heat-on-the-surface family as Automatic Weights, with a more careful solver that bound every mesh the bone-heat version dropped. The Voxel Heat Diffuse page<br>benchmarks it next to the voxel solver.
Automatic Weights is the fastest when it runs , but it left 44 of 102 Mixamo characters and 15 of 100 RigNet with no weights at all, which is why its all-inclusive average sits so low. On the characters it does bind it reaches about 68% pp. The failures are the dense, multi-part, and non-manifold meshes its surface formulation cannot solve.
Conclusion
On this benchmark the volumetric binder tops every quality column and edges out Voxel Heat Diffuse while running 3 to 250 times faster, with Surface Heat Diffuse close behind on Mixamo and the fastest of the diffuse methods. The number to watch as the binder matures is whether it holds that lead on geometry outside the test set.
Methodology
The datasets and exact metric definitions, kept here as a reference for the numbers above.
The datasets
Two public character sets, each shipping the artist’s own skin weights as the ground truth:
Mixamo (mixamo.com<br>), Adobe’s library of rigged production humanoids. The 102 characters here are the full skinning-test set, from clean basemeshes to clothed, multi-part...