viblock5 — Training Results

June 24, 2026

run je8e2b2u — 2026-06-21 to 2026-06-22 — 6 A100s — 100 epochs

viblock5 is a clean re-baseline of viblock4: SigLIP-B/16 with 8³ voxel slot tokens injected at layer 10, decoded to 32³ via 3D transposed convolutions, with visibility-masked loss and audit bugs fixed. Loss is applied only to visible (vis=2) and previously-seen (vis=1) voxels; never-seen cells are masked out entirely.

Training complete. WandB run →

vis occ_iou (epoch 100) 0.9878 ↑ from viblock4: 0.7635
type_acc (epoch 100) 0.8405 ↑ from viblock4: 0.7226
val_loss (epoch 100) 0.0021
full occ_iou (epoch 100) 0.2545 includes never-seen voxels

Loss Curves

viblock5 training curves: vis occ_iou, type_acc, val_loss over 100 epochs

Dashed line on occ_iou = viblock4 best (0.7635). Eval every 5 epochs.

Epoch 100 Renders

Each image is a triptych: raw game frame → ground truth (Furnace GPU render) → model prediction.

validation samples

val sample 0 epoch 100
val 0
val sample 1 epoch 100
val 1
val sample 2 epoch 100
val 2

train samples

train sample 0 epoch 100
train 0
train sample 1 epoch 100
train 1
train sample 2 epoch 100
train 2