artokun/nic0le-krea2

nic0le — Krea 2 LoRA

Character LoRA for Krea 2 Raw. Trigger word: nic0le.

Trained with ostris/ai-toolkit.

Checkpoints

File	Steps
nic0le_000002000.safetensors	2000
nic0le_000002250.safetensors	2250
nic0le_000002500.safetensors	2500
nic0le_000002750.safetensors	2750
nic0le.safetensors	3000 (final)

Identity holds from roughly step 2000 onward. Earlier checkpoints are included so you can pick your own point on the likeness-vs-flexibility curve.

Training

Base	krea/Krea-2-Raw (arch krea2, qfloat8 quantized for training)
Network	LoRA, linear 32 / alpha 32
Steps	3000, batch size 1
Optimizer	adamw8bit, lr 1e-4, constant
Scheduler	flowmatch, linear timesteps
Dataset	19 images at 1024px, captioned with nic0le as the leading token

Usage in ComfyUI

UNETLoader   krea2_raw_bf16.safetensors        → LoraLoaderModelOnly (this LoRA) → KSampler
CLIPLoader   qwen3vl_4b_bf16.safetensors  type=krea2
VAELoader    qwen_image_vae.safetensors

Sampler settings for raw (not turbo): ~40 steps, cfg 3.0–3.5, euler / simple. Turbo is the distilled variant and wants ~8 steps at cfg 1 instead.

Use the bf16 text encoder, not fp8

qwen3vl_4b_fp8_scaled.safetensors produced uniform black images on any descriptive prompt in our testing — no error raised, full step count, black output. Single-token prompts survived numerically but carried no usable signal, so results were driven by the seed alone. Switching to qwen3vl_4b_bf16.safetensors (from Comfy-Org/Krea-2) fixed it completely.

If you are chaining Krea 2 with a second model family in one workflow, free VRAM between model loads. Loading Krea 2 alongside another large model corrupted the second model's weights silently — sampling to NaN with no error in the log.