nala-cs-krea2-lora

nala-cs-krea2-lora

Krea 2 character LoRA — trigger n4la woman. Trained on Krea 2 RAW, intended for inference on Krea 2 Turbo.

Status: ✅ training complete

Training configuration

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.

min avr_loss 0.0382 · final 0.0389

Inference (Krea 2 Turbo, Musubi Tuner)

Prepend the trigger n4la woman to the prompt. Recommended: stack over the house style LoRA.

python src/musubi_tuner/krea2_generate_image.py \
    "n4la 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:

Nala_cultureCS-portrait-001-seedream.jpg

n4la woman, close-up shot, one hand raised near her face with fingers loosely curled, wearing a cream tank top with thin straps, a red-yellow-green patterned headband and gold hoop earrings, a wide joyful smile showing teeth, plain white background.

Nala_cultureCS-medium-011-seedream.jpg

n4la woman, medium shot, fingertips resting lightly against her own chin near her parted lips, wearing a beige halter crop top with a deep plunging neckline and beige high-waisted pants with a brown belt, a red-yellow-green patterned headband and gold hoop earrings, a soft coy smile, plain white background.

---

Auto-generated by training watchdog.