elsa-krea2-lora

Krea 2 character LoRA — trigger 3lsa woman. Trained on Krea 2 RAW, intended for inference on Krea 2 Turbo.

Status: ✅ training complete

Training configuration

Field	Value
Base model (train)	krea/Krea-2-Raw (raw.safetensors, 12B DiT, undistilled)
Inference target	krea/Krea-2-Turbo (8-step distilled) — RAW-train / Turbo-infer
VAE	Qwen-Image VAE (qwen_image_vae.safetensors)
Text encoder	Qwen3-VL-4B (qwen3vl_4b_bf16.safetensors)
Trainer	Musubi Tuner (kohya-ss), networks.lora_krea2
LoRA	rank/dim 24, alpha 24, all Linear layers of the DiT
Optimizer	adamw8bit, lr 1e-4
Precision	bf16, gradient checkpointing, SDPA
Timestep sampling	krea2_shift (resolution-aware), weighting_scheme none
Trigger	3lsa woman (skin + blonde hair (updo) + green eyes + face baked into the trigger; black-framed glasses + gold pendant + clothing described)
Dataset	56 images + NL captions (Seedream, Western-animation style); black-framed glasses + gold pendant + light-blue button-up shirt + white high-waisted pants + pose/expression/framing/background described, identity in trigger. batch 1, 1024 multi-res buckets, num_repeats 2
Schedule	12 epochs = 1344 steps, save every 2 epochs, seed 42
Hardware	1x NVIDIA H100 80GB (RunPod, EU-NL-1, Musubi volume)

Loss per checkpoint

avr_loss is a rolling average — flow-matching loss is near-flat, so use it as a sanity check and pick the best checkpoint visually.

checkpoint	epoch	step	avr_loss
-000002	2	224	0.0529
-000004	4	448	0.0472
-000006	6	672	0.0458
-000008	8	896	0.0469
-000010	10	1120	0.0503
final	12	1344	0.0439

min avr_loss 0.0412 · final 0.0439

Inference (Krea 2 Turbo, Musubi Tuner)

Prepend the trigger 3lsa woman to the prompt. Recommended: stack over the house style LoRA.

python src/musubi_tuner/krea2_generate_image.py \
    "3lsa woman, <scene>" \
    --dit turbo.safetensors --vae qwen_image_vae.safetensors \
    --text_encoder qwen3vl_4b_bf16.safetensors \
    --steps 8 --guidance_scale 1 --mu 1.15 --width 1024 --height 1280 \
    --attn_mode torch --lora_weight <checkpoint>.safetensors --lora_multiplier 1.0

Dataset samples

Representative image+caption pairs from the training set:

Elsa-portrait-001-seedream.jpg

3lsa woman, close-up portrait, facing the camera, wearing black-framed glasses, a gold pendant necklace, and a light blue button-up shirt with the collar open, a subtle warm smile, plain off-white background.

Elsa-medium-001-seedream.jpg

3lsa woman, medium shot, one hand resting on her hip, wearing black-framed glasses and a gold pendant necklace, a sheer light blue button-up shirt unbuttoned low and white high-waisted pants with a brown belt, lips slightly parted in a sultry expression, plain light purple background.

---

Auto-generated by training watchdog.