millie-v4-krea2-lora

Krea 2 character LoRA — trigger 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	woman (no dedicated trigger token — class word only; identity learned from the dataset (young woman))
Dataset	66 seedream-v5-pro 2k + a few gpt-image Western-animation seductive portrait+medium images + paired natural-language captions; clothing/pose/expression/framing/background described; young woman, no dedicated trigger (class word 'woman'). batch 1, 1024 multi-res buckets, num_repeats 2
Schedule	12 epochs = 1584 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	264	0.0543
-000004	4	528	0.0492
-000006	6	792	0.0465
-000008	8	1056	0.0480
-000010	10	1320	0.0472
final	12	1584	0.0467

min avr_loss 0.0446 · final 0.0467

Inference (Krea 2 Turbo, Musubi Tuner)

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

python src/musubi_tuner/krea2_generate_image.py \
    "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:

Millie-portrait-001-seedream.jpg

woman, close-up portrait facing the camera with her head slightly tilted, no hands visible, a soft closed-lip smile and a relaxed gaze, wearing a white scoop-neck top just visible at one shoulder, plain off-white background.

Millie-medium-001-seedream.jpg

woman, medium shot, standing and facing the camera with arms relaxed at her sides and hands near her thighs, wearing a white long-sleeve tie-front crop top with a deep V neckline over a beige mini skirt with a side slit, a calm expression with lips slightly parted, plain light purple background.

---

Auto-generated by training watchdog.