udbbdh/toposet-rig-vae-z16

TopoSet rig VAE, z16 / KL 1e-4 (RigSpacetimeBranchVAE) — intermediate checkpoint

Weights only. 16-channel per-token latent VAE over [vertices | joints] tokens of a rigged low-poly mesh (vertex coordinates via a Huber regression head, edges/faces/orientation, bone parents, skin weights; no decoder RoPE, decodes with coordinates unknown). Initialised from the z32 model by PCA-projecting the bottleneck onto the top-16 latent directions (95.3% of variance), then continued with KL 1e-4 (unified normalisation).

• rig_vae_ema_step_30000.pth — EMA (0.999) weights at step 30000 of a 300000-step run; still training, the topology heads are still recovering from the bottleneck cut (see below)
• config.yaml — training config; models.vae.args are the constructor arguments (z_dim: 16)

Validation at this checkpoint (FLR held-out val, EMA weights)

step	edge F1	face P/R	orientation acc	vertex MAE	bone parent acc	skin L1
30000	0.965	1.000 / 0.9999	0.985	0.00103 (grid 1/1024 = 0.00098)	1.000	0.437

For reference the z32 parent reaches edge F1 1.000, vertex MAE 0.00091, skin L1 0.225. Code is not included.