From SDXL to Krea2: Can Stacked LoRAs Be Replaced by Full Model Finetuning?
Bitter_Bag_3429 · reddit · 2026-10-28
A developer shares his LoRA training workflow from SDXL to Krea2, having used Kohya, Onetrainer, and fizgig (which he considers the best), and found RunPod rental fast and cheap — about an hour of training per LoRA at modest cost.
He now asks whether it's feasible to finetune the full Krea2 base model on 20+ prepared, captioned, and masked datasets at once, to avoid conflicts between stacked LoRAs, and wants real-world figures: GPU/VRAM/storage requirements, training hours, and total cost.
More from Infra
- SK hynix down 48% as DDR5 spot jumps 31%: market bets the memory boom ends — tengyanAI · 2026-10-28
- fastconstmap 0.11.0 adds free-threaded Python support at ~9 bytes per key — lemire · 2026-10-28
- Two days tuning one Tesla P40: 35B MoE with full 262k context at ~60 t/s — sargetun123 · 2026-10-28
- RunPod cold starts hit 324s median for 32B models — one team's GPU alternative comparison — Kind_Taste_3985 · 2026-10-28
- NVIDIA launches DSX AI factory designs: near-zero cooling water, 30% less power — RevLebaredian · 2026-10-28
- TigerBeetle creator on infra code: 'The database is the last bastion of software craft' — arpit_bhayani · 2026-10-28