Fixing MOSS LoRA training: clean-end sigma share hits 68% likeness at epoch 12 vs 67% at 30
shootthesound · reddit · 2026-10-19
A hands-on deep dive into why MOSS LoRA training struggles with likeness: its native sigma schedule allocates only 11% of steps to the clean end (sigma<0.5) where faces and fine detail are refined, versus 18-28% for Krea 2. The author shipped Fizgig v8.7.0 with a new Clean-end share option (default 60%).
On the same 53-image dataset, rank 16, LR 1e-4, same seed, measured by ArcFace similarity:
- Old schedule: best 67% at epoch 30
- Clean-end 60%: 68% at epoch 12 (640 steps vs 1,590), stable at 68-69% instead of swinging 10 points per epoch
The author stresses 60 is experimental and invites the community to test other values (30, near Krea 2's level, is a sensible next try). Tool and release notes on GitHub (Fizgig).
More from coding & agent
- Every engineering leader is building the same AI automations right now — generativist · 2026-10-19
- 'Prayer-based programming' emerges as AI coding's latest self-deprecating meme — lxfater · 2026-10-19
- Developer: I keep playing 'human API' in my AI workflows — jacalulu · 2026-10-19
- Vibe coding a parent's TV: Claude Code strips 25 Xiaomi packages, kills boot ads — HarperSCarroll · 2026-10-19
- Sentry co-founder's AI writing workflow: dictate ideas, let an agent outline, write it yourself — lmoroney · 2026-10-19
- CarryOn: object-oriented context runtime for long-horizon agents boosts OSWorld by 5.8% — McAuley-Lab · 2026-10-19