Appreciate it! Correct — r=128 (alpha=64). Settings are pretty standard: AdamW 8-bit with a constant LR schedule; I slightly lowered the LR at each stage based on results. Instead of a single long run, I did 15 short continuation passes, ~1k images per stage with changing subsets (some overlap; ~15k overall). In practice the consistency is driven by the large, curated data and the staged training, more than by any special parameter tweaks.

I realize 15k images is a lot for a single subject; the transferable part is the staged continuation on curated subsets with a conservative LR and early stops — that tends to preserve consistency more than the exact LR numbers.