← ImageLoRA · All LoRA guides
Krea 2 LoRA training with musubi-tuner: setup guide
Setting up musubi-tuner to train a Krea 2 LoRA: support status, the settings known to work, and links to the docs. Status in musubi-tuner: supported (fp8 / ConvRot int8 base).
About musubi-tuner
kohya's CLI scripts for newer models; supports fp8 base weights and recommends 12 GB+ VRAM for image training. Sources: github.com/kohya-ss/musubi-tuner
VRAM
ai-toolkit: ~20 GB at batch 1 with qfloat8; also trains on 12 GB with Krea Raw, low VRAM on, 50% transformer offload, 100% text-encoder offload, convrot 8, compile off, LoKr factor 8, 512 resolution only, and cached latents plus text embeddings. musubi: 12GB+ recommended for image training (general guidance); fp8 or --convrot_int8 halves weight VRAM
Settings that work for Krea 2 (any trainer)
- 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
- Training at 512px can work for output up to ~2MP; add higher-res buckets gradually
- ai-toolkit downscales images larger than 1536px
- Automagic3 can give weaker LoRAs than v1/v2; pair it with LoKr factor 8 or 4 and sigmoid timesteps (that's my recipe). ai-toolkit's experimental automagic4 (Oct 23) keeps the LR from drifting to the floor
- EMA in ai-toolkit makes no noticeable difference; Automagic v3 already behaves a lot like EMA
- OneTrainer preview-2 branch (not master yet): LoRA on 8-bit quantized Krea 2 trains ~1.2x faster (16 GB preset); FLUX.2 16 GB preset ~1.3x; faster 8-bit math on RTX 50-series; DoRA not sped up
- Character captions: name, expression, pose, clothing, then scene
- Kaloscope style adapters (heathcliff01) for Krea 2 Raw and Turbo: style by artist name, weighted blends or a reference image; needs its own ComfyUI nodes
- ai-toolkit: point to local or alternative Qwen3-VL text encoder and Qwen-Image VAE files with text_encoder_path and vae_path under model_kwargs in the YAML
- Very slow steps (hundreds of seconds per step) mean the GPU is out of VRAM and swapping; lower the batch size or turn on low VRAM / offloading
- Pose control for Krea 2 Turbo: the Krea-2-PoseControl LoRA (MIT, 17-keypoint skeleton) with the ComfyUI-Krea2-Pose node pack; use release/krea2-pose-control-mix025.safetensors
- 12 GB ai-toolkit recipe (Krea Raw): low VRAM on, 50% transformer offload, 100% text-encoder offload, convrot 8 quantization, compile off, LoKr factor 8, 512 resolution only, cache latents and text embeddings. First steps can take many minutes; if the run hangs after the first step, try dropping the training adapter, a lower quant (Q6/Q4), or skipping text-encoder offload once embeddings are cached
General tips
- Gated Hugging Face models: click 'Agree and access' on each model page, or you'll get 403 errors
- Steps scale with dataset size (steps = epochs x images at batch 1)
- A safe start: LoKr with auto rank (~64), batch/grad accumulation 4
- Leave out of your captions whatever you want burned into the LoRA; caption what should stay changeable
- ai-toolkit dataset folder weights let you train the same images with long, medium and short captions in separate folders
- ai-toolkit supports validation loss, which shows when a run stops improving
- In ai-toolkit, Cache text embeddings works for basically every model and frees the text encoder from VRAM and RAM
- Prefer int8 convrot over qfloat8 quantization in ai-toolkit: better quality at the same memory
- Gradient accumulation and batch size don't scale the learning rate in ai-toolkit; a bigger batch needs about 1.5-2.5x the steps
Compare with other trainers
Where to rent a 16 GB+ GPU
Cheapest 16 GB+ cards (2026-10-20 08:06 UTC) from RunPod and Vast.ai. Prices move constantly; see the live price table.
| GPU | VRAM | RunPod community $/hr | Vast.ai cheapest $/hr |
|---|
| Tesla V100 | 16 GB | $0.19 | $0.08 |
| RTX 3090 | 24 GB | $0.22 | $0.12 |
| RTX 5060 Ti | 16 GB | — | $0.14 |
| Tesla V100S | 32 GB | — | $0.14 |
| RTX A5000 | 24 GB | $0.16 | — |
| RTX A4000 | 16 GB | $0.17 | — |
Sources
FAQ
Does musubi-tuner support Krea 2?
Yes: supported (fp8 / ConvRot int8 base) (see sources below).
How much VRAM do I need for a Krea 2 LoRA?
ai-toolkit: ~20 GB at batch 1 with qfloat8; also trains on 12 GB with Krea Raw, low VRAM on, 50% transformer offload, 100% text-encoder offload, convrot 8, compile off, LoKr factor 8, 512 resolution only, and cached latents plus text embeddings. musubi: 12GB+ recommended for image training (general guidance); fp8 or --convrot_int8 halves weight VRAM
Weekly newsletter: new models, trainer releases, cheapest GPUs.
Subscribe
Data updated 2026-10-20. Disclosure: RunPod and Vast.ai links are referral links; ImageLoRA may earn a commission at no cost to you. unknown = no data yet. · Raw data (JSON) · llms.txt