kalaya-krea2-lora

Krea 2 character LoRA — trigger k4ly woman. Trained on Krea 2 RAW, for inference on Krea 2 Turbo.

Status: training complete

Field	Value
Base model	krea/Krea-2-Raw (raw.safetensors, 12B DiT)
Trainer	Musubi Tuner, networks.lora_krea2
LoRA	rank/dim 24, alpha 24
Optimizer	adamw8bit, lr 1e-4
Precision	bf16, SDPA
Timestep sampling	krea2_shift, weighting_scheme none
Trigger	k4ly woman
Dataset	57 Seedream-v5-flash 2k frames (31 portrait + 24 medium + 2 full-body), chat set: flirty/seductive expressions, every frame in a different outfit (none in her reference outfit), plain hex-colour backgrounds. Natural-prose captions. Messy updo and gold bracelets are baked into the trigger; gold hoop earrings and a thin gold necklace are captioned where visible. batch 1, 1024 buckets, num_repeats 2
Schedule	12 epochs = 1368 steps, save every 2 epochs, seed 42
Econ run	no --gradient_checkpointing + PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

Loss per checkpoint

avr_loss is a rolling average and near-flat for flow-matching — pick the checkpoint visually.

checkpoint	epoch	avr_loss
-000002	2	0.048
-000004	4	0.0426
-000006	6	0.0415
-000008	8	0.0403
-000010	10	0.0424
final	12	0.037

Inference

Prepend k4ly woman. Recommended: stack over the house style LoRA.