Slow Krea2 LoRA training on RTX 5080 with Fizgig: full config shared, asking for tuning help
BeautifulStation4 · reddit · 2026-11-09
A LoRA training beginner posted for help: training a Krea 2 model (bf16) with Fizgig on an RTX 5080 is extremely slow, and he asked what config parameters others use. He shared his full setup: 42 epochs / 2814 steps, batch 1, 1.0 MP dataset, adamw8bit optimizer (lr 0.0001), LoRA rank 32 / alpha 32 with EMA 0.98, block swap 8, constant LR scheduler, plus caching, bucketing, and per-epoch sampling/saving. The value is in experienced users diagnosing this config in the comments.
More from Multimodal
- Hy-Nomas + Hy4 pipeline turns imagined world-model frames into playable three.js games — ShunyuYao12 · 2026-11-09
- Grainy, blurry AI images aren't "realism" — clarity should be the bar — Deep_Imagination_811 · 2026-11-09
- AI-generated anime battle scene: swordsman vs crystal monsters — Omega_Criative · 2026-11-09
- AI biweekly recap #103: 351 news items on Dior, Prada, Publicis, Gemini, Kling, Canva — emmanuelvivier · 2026-11-09
- PlayCanvas open-sources Matrix-style digital rain shader for 3D Gaussian splats — willeastcott · 2026-11-09
- Reddit user asks: how to train a LoRA that stays compatible with other LoRAs — jonbristow · 2026-11-09