Skip to content

Latest commit

 

History

History
executable file
·
86 lines (66 loc) · 3.72 KB

File metadata and controls

executable file
·
86 lines (66 loc) · 3.72 KB

Deformation Activation Audit

Result

Exact upstream SC-GS hard deformation activation failed on the controlled procedural scene. The primary mechanism was the opacity reset scheduled at the same update as hard deformation activation. The optimizer step then added a secondary deformation overshoot.

This is a negative upstream result. It is not relabeled as a successful unmodified SC-GS baseline.

Machine source: /mnt/d/Track4D-GS-outputs/activation_audit/activation_abcd_run2_node/summary.json.

Controlled boundary decomposition

The audit restored the accepted update-2,999 checkpoint in a fresh process and used camera r_000, time 0.0, the restored RNG state, and the same active SH degree for all renders.

State Meaning PSNR SSIM alpha mean visible
A pre-activation detached forward 14.6476 0.57042 0.99372 278,636
B attached full-deformation forward, pre-backward 14.6476 0.57042 0.99372 278,636
C4 post-backward and post-densification, pre-reset 13.9804 0.51011 0.99309 289,011
C5 post-opacity-reset, pre-optimizer 8.8589 0.12350 0.13983 289,011
D post-optimizer 8.9064 0.11209 0.25577 289,660

A and B were tensor-identical for RGB, alpha, canonical positions, translation, rotation, scaling, and transformed positions. B and the post-backward state were also tensor-identical. Therefore neither attaching autograd nor backward changed the forward state.

Densification caused a moderate quality change. The C4-to-C5 opacity reset changed no geometry or deformation outputs, but changed RGB by up to 1.67157 and alpha by up to 0.99003; this is the primary discontinuity. The C5-to-D optimizer step then changed translation/transformed position by up to 1.45827, rotation output by 0.06750, and scaling output by 0.06569.

Pinned stage semantics

At the update-3,000 boundary the pinned source:

  • attaches the deformation network and control nodes after the warm-up;
  • activates translation, quaternion rotation, and additive scaling together;
  • applies the local-frame control-node warp before raster projection;
  • executes Gaussian densification and the scheduled opacity reset;
  • keeps deformation-network and control-node parameters in their existing optimizer groups and learning-rate schedules;
  • does not perform node densification at this boundary.

The failure was therefore classified as a coincident maintenance discontinuity with a secondary optimizer contribution, not as a camera or coordinate error.

Stabilization decision

The selected named protocol is Stabilized SC-GS, distinct from exact upstream SC-GS:

  • a 1,000-update smoothstep coefficient;
  • translation alpha * t_predicted;
  • identity-to-predicted quaternion SLERP for rotation;
  • scaling interpolated in the additive representation consumed by SC-GS;
  • only the opacity reset coincident with activation is skipped;
  • original optimizer learning rates;
  • no coordinate changes, position clamps, motion truncation, LR reduction, or gradient clipping.

The 500-update bridge failed after three focused engineering repairs: a stale reference, zero-angle SLERP NaNs, and then non-finite rotation at update 3,200. The 1,000-update bridge passed 300 full-deformation updates and was extended to 500 full-deformation updates.

Accepted checkpoint:

/mnt/d/Track4D-GS-outputs/stabilized_scgs/stabilized_scgs_attempt2_full_extension200_node/checkpoints/main_update_04499.pt

  • SHA-256: 9e5f63cd643463183fa5f5e725ee4f6ff38331a3b8ef82db28f89a179fbf8216
  • size: 387,945,298 bytes
  • Gaussians: 444,687
  • control nodes: 128
  • peak reserved VRAM: 4,649,385,984 bytes
  • alpha: exactly 1
  • fully activated updates: 500
  • fresh-process restore and one additional update: passed