krea2-style-v4

Krea 2 style LoRA (digital art). Trained on Krea 2 RAW, intended for inference on Krea 2 Turbo.

Status: 🟢 training in progress — step 4120/4120

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-Instruct (qwen3vl_4b_bf16.safetensors)
Trainer	Musubi Tuner (kohya-ss), networks.lora_krea2
LoRA	rank/dim 32, alpha 32, all 264 Linear layers of the DiT
Optimizer	adamw8bit, lr 1e-4
Precision	bf16, gradient checkpointing, SDPA
Timestep sampling	krea2_shift (resolution-aware), weighting_scheme none
Dataset	348 images + paired captions (combined: 249 style/SFW + 19 male body-type + 13 nude/flaccid-penis + 67 Genatomy anatomy close-ups, liquids mostly removed), batch 4, 1024 multi-res buckets, num_repeats 1
Schedule	40 epochs = 4120 steps, save every 5 epochs, seed 42
Hardware	1x NVIDIA H100 80GB (RunPod)

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
-000005	5	515	0.0732
-000010	10	1030	0.0727
-000015	15	1545	0.0699
-000020	20	2060	0.0730
-000025	25	2575	0.0720
-000030	30	3090	0.0689
-000035	35	3605	0.0692

min avr_loss 0.0656 · final 0.0701

Inference (Krea 2 Turbo, Musubi Tuner)

python src/musubi_tuner/krea2_generate_image.py \
    "your prompt" \
    --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 1024 \
    --attn_mode torch --lora_weight <checkpoint>.safetensors --lora_multiplier 1.0

Dataset samples

Two representative image+caption pairs from the 348-image training set:

z61flac_13.jpg

A medium shot of a muscular, naked man sitting on a black leather bench in a locker room, displaying his flaccid penis and glans, with wooden lockers in the background.

g0000-raw.jpg

A close-up, low angle view shows a woman's vulva and labia with a patch of dark pubic hair above. Her legs are spread wide, with her skin bathed in a warm, reddish light.

---

Auto-generated by training watchdog.