I audited deformation locality in pretrained HUGS digital humans. The question was not whether the model could render a person, but whether a small articulated edit stayed anatomically local once motion passed through learned Gaussian skinning weights.
This is a diagnostic study of an existing method. It does not claim a new correction architecture. The current public result is intentionally bounded to the tested checkpoints, joints, and poses.
A wrist or elbow edit naturally propagates through its kinematic descendants. The useful locality question is whether Gaussians anatomically assigned to the opposite upper body acquire learned support from the perturbed branch, and whether those small weights causally mediate the resulting response.
The left wrist was the primary probe. The left elbow was selected and frozen before inspecting elbow displacement: joint 18, z-axis, +10 degrees, with branch channels [18,20,22]. Failures were kept as results rather than changing the endpoint after inspection.
Independently pretrained HUGS NeuMan models: Seattle, Parkinglot, and Jogging.
Subject-specific K6 replacement and selective removal of learned perturbed-branch LBS components on the fixed contralateral subset.
Seattle 4 poses, Parkinglot 4 poses, Jogging 10 poses, for 18 nested tests per joint.
All six joint × checkpoint diagnostic cells passed the frozen criteria. The independent model-level unit remains the checkpoint, n=3. Joint and pose observations are repeated diagnostics within those models.
| Checkpoint | Joint | Poses | Min K6 reduction | Min ablation reduction | Min mass/reduction corr | Max ablated contra |
|---|---|---|---|---|---|---|
| Seattle | Left wrist | 4 | 99.8175% | 100% | 0.9808 | 0.0 |
| Seattle | Left elbow | 4 | 99.8187% | 100% | 0.9689 | 0.0 |
| Parkinglot | Left wrist | 4 | 99.7086% | 100% | 0.9975 | 0.0 |
| Parkinglot | Left elbow | 4 | 99.8913% | 100% | 0.9959 | 0.0 |
| Jogging | Left wrist | 10 | 99.8178% | 100% | 0.9899 | 0.0 |
| Jogging | Left elbow | 10 | 99.8881% | 100% | 0.9846 | 0.0 |
Combined benchmark: 3 independently pretrained checkpoints, 2 tested joints, 36 nested pose diagnostics, 6/6 joint × checkpoint cells passed. Global minimum K6 reduction was 99.708575%; selective branch ablation removed 100% of the tested contralateral response in every nested pose diagnostic; global minimum removed-mass correlation was 0.968898.
Several provenance and implementation assumptions were contradicted during the audit. They remain part of the research record because they change how the evidence should be interpreted and reproduced.
A Seattle count mismatch was initially suspected to come from ≥0.9 versus >0.9. The audit disproved that and traced the difference to the historical saved anatomy source.
The downloaded configs said hugs_triplane, so I tested whether an alias mapped it to HUGS_TRIMLP. No such alias exists in the released source. Exact numeric reconstruction still matched the checkpoint tensors.
An early ankle interpretation counted expected same-leg upstream propagation as leakage. The target was corrected to separate local kinematic propagation from the nonlocal remainder.
A full-pose cell was once truncated during transfer and failed with a syntax error before scientific execution. It was logged separately rather than confused with an experimental failure.
For articulated 4D representations, controllability is also a representation property. Small unintended cross-joint weights can produce structured motion far from the edited body part. Diagnosing that behavior gives me a concrete bridge from Gaussian representation research into interactive digital humans and controllable 4D media.