Training MOSS face LoKr on RTX 5060 plateaus at ~60% likeness after 100+ epochs
Confident_Ad2351 · reddit · 2026-11-03
A hobbyist with SDXL LoRA experience tried training a LoKr for the MOSS image model on a 16GB RTX 5060 to reproduce a movie star's face, with poor results: 40 training images, bf16 base model, Automatic v3 optimizer, 100+ epochs, but likeness plateaus around 60% — outputs look like "the star's sister."
He used the newly discovered ZenTrainer by Nynxz (nice interface), with no training errors, and doesn't blame the tool since the same images worked for SDXL. He's asking the community for their face/body reproduction recipes.
More from Multimodal
- Claymation-style blue-and-pink AI-generated botanical portraits go viral — gizakdag · 2026-11-03
- Grok explains the etymology of 'kumi-daiko' while composing taiko music — TinfoilTricorn · 2026-11-03
- GANPaint, the classic mechanistic interpretability tool, revived via vibe coding — hen_str · 2026-11-03
- AI-generated 'Nike Shades' concept with swoosh devil horns is absurdly convincing — Delahuntagram · 2026-11-03
- GLB-3D-Coder: a tiny 4B model generates game-ready 3D assets with GLB export — canalCCore2 · 2026-11-03
- AI artwork 'Digital World' lands Midjourney Magazine Issue 42 poster — ciguleva · 2026-11-03