How to train a Krea 2 LoRA

Which trainers support Krea 2 / Krea 2 Turbo, how much VRAM you need, and settings that are known to work. Every row links to its source.

Trainer support

Trainer	Status
ai-toolkit	supported (Raw + Turbo, incl. edit training)
OneTrainer	supported (per community blog; not in README list)
musubi-tuner	supported (fp8 / ConvRot int8 base)
kohya sd-scripts	not listed
SimpleTuner	supported

VRAM

Not published by trainers. musubi: 12GB+ recommended for image training (general guidance); fp8 or --convrot_int8 halves weight VRAM

Known-good settings

Natural-language captions, ~70-500 words; shorter captions lock features harder
Diverse dataset (varied expressions/angles)
Train at 1024px or higher
Batch size >= 2 if VRAM allows

My proven Krea 2 recipe

Trainer	ai-toolkit
Base model	krea/Krea-2-Raw (arch krea2)
Network	LoKr: linear 32, alpha 32, full rank on, factor 8
Optimizer / LR	automagic3, 1e-4
Weight decay	1e-5 (0.00001)
Schedule	flowmatch scheduler, sigmoid timesteps, content_or_style balanced
Differential guidance	On, scale 3
Batch	1, grad accumulation 1
Memory	gradient checkpointing, bf16, qfloat8 model + text encoder, low_vram, cached latents and text embeddings
Resolution	1024
Captions	dropout 0.05; text encoder not trained; trigger word optional
Epochs	About 160 epochs (my best checkpoint: ~162 epochs on a 34-image dataset; batch 1, repeats 1, grad accumulation 1)
Samples	guidance 4, 30 steps
Note	Results may vary; the right epoch count depends on your dataset size and content.

# Example ai-toolkit config (generic hyperparameters only)
config:
  process:
    - type: sd_trainer
      network:
        type: lokr
        linear: 32
        linear_alpha: 32
        lokr_full_rank: true
        lokr_factor: 8
      train:
        batch_size: 1
        gradient_accumulation_steps: 1
        steps: 5500            # steps = epochs x images (e.g. 162 x 34 ≈ 5,500); ~160 epochs worked best for me
        train_text_encoder: false
        gradient_checkpointing: true
        noise_scheduler: flowmatch
        timestep_type: sigmoid
        content_or_style: balanced
        optimizer: automagic3
        lr: 1e-4
        optimizer_params:
          weight_decay: 1e-5   # 0.00001
        dtype: bf16
        cache_text_embeddings: true
        do_differential_guidance: true
        differential_guidance_scale: 3
      datasets:
        - caption_dropout_rate: 0.05
          cache_latents_to_disk: true
          resolution: [1024]
      model:
        name_or_path: krea/Krea-2-Raw
        arch: krea2
        quantize: true
        qtype: qfloat8
        quantize_te: true
        qtype_te: qfloat8
        low_vram: true
      sample:
        guidance_scale: 4
        sample_steps: 30
      # trigger_word: optional / not needed

Where to rent a GPU

See cheapest GPU to train a LoRA (RunPod, Vast.ai).

Sources

Weekly newsletter: new models, trainer releases, cheapest GPUs. Subscribe

Data updated 2026-10-16. Some provider links may be referral links. unknown = not yet verified.