#13, 21.790

@will

3

0

Run #13 used the exact same recipe as Run #12 (cap=4M, 50K steps, bilateral grid for per-view exposure, DA3-Streaming depth at λ=0.02) plus one additional regularizer: PGSR-style multi-view consistency loss (--multi-view-weight 0.05 --multi-view-from-iter 7000), which samples patches from nearby v