ariaof_krea2_v2

Character LoRA (LoKr) for Krea 2 Raw. Prompt with ariaof.

Which file to use

File	Step	Epoch	Notes
aug2_ariaof_v2_5000.safetensors	5000	32	latest, longest trained
aug2_ariaof_v2_4000.safetensors	4000	25	mid-point, use if 5000 overfits
ariaof_krea2_v2.safetensors	3000	19	the original 2026-08-01 run

The aug2_ files are the 2026-08-02 continuation of the same run — resumed from step 3000 with its optimizer state, so they are one continuous 5000-step training, not separate runs.

Naming note: do not confuse these with ariaof_raw_2_4000 / ariaof_raw_2_5000, which are a different, earlier LoRA (diobrando0/ariaof-krea-raw-2) that happens to share the step numbers. The aug2_ prefix exists to keep them apart.

Training

base	krea/Krea-2-Raw (gated), arch krea2, quantized qfloat8 incl. TE
network	LoKr — linear 32, linear_alpha 32, lokr_full_rank: true
optimizer	adamw, weight_decay 1e-4, lr 1e-4
scheduler	flowmatch, timestep_type: sigmoid, loss mse
resolution	1280 / 1024 / 1328 buckets, batch 1, bf16
dataset	50 images + captions, caption dropout 0.05
trainer	ai-toolkit 0.12.0
final loss	4.911e-02

Krea 2 Raw needs the turbo LoRA (krea2_turbo_lora_rank_64_bf16) at strength >= 0.75 — 0.85 recommended — or it will not converge at the low step counts these graphs use.

Full training state (optimizer + config + dataset) for resuming further: diobrando0/ariaof_krea2_v2_training (private).