Krea2 LoRA Training Confusion: Loss Graphs Show Best at Low Steps but Actual Results Underfit
jarrodthebobo · reddit · 2026-07-24
A Reddit user shares confusion training character LoRAs with Krea2: validation/loss graphs indicate best at 1000-1500 steps, but actual testing shows better likeness at 4000-5000 steps. Dataset 100 images, tried various parameters, seeks objective method to select best LoRA.
More from Research
- NeurIPS Peer Review is Highly Random: 50% of Spotlights Were Rejected in RCT — NeelNanda5 · 2026-07-24
- Google Releases ATLAS Report: Reveals Real AI Usage Based on 15M Interactions — soumitrashukla9 · 2026-07-24
- Nature paper finds GPT-4 can predict social science experiment results with r=0.85 — JeremyNguyenPhD · 2026-07-24
- Qwen MoE stability paper shows training-inference KL divergence rising over steps — heghbalz · 2026-07-24
- AI math breakthroughs keep proving earlier limits wrong, researcher says — herbiebradley · 2026-07-24
- New Aggregation Methods Upgrade Permutation Tests with Uniform Power Dominance — ArthurGretton · 2026-07-24