How to train a LoRA for Krea 2?

last edited

A Krea 2 LoRA is trained in the hosted Krea Train tool with a minimum of 3 captioned images. The other option is training locally with ai-toolkit and the gated Krea 2 Raw weights. The hosted beta started in May 2026 for Max and Business subscribers.

Which route trains a Krea 2 LoRA?

routeneedsdatasetsteps and rank
hosted trainerMax or Business plan, no GPU setup3+ images, captions requireddefault steps, rank not exposed
ai-toolkitgated weights, 24 GB card typical10-50 images with .txt captions1000-3000 steps, rank 32
gui wrapper24 GB VRAM, 32 GB RAM, 100 GB disk10-50 images, auto-tagging2000 steps, rank 32, alpha 32

The hosted route has 3 steps: open /train, add images, pick Krea 2 Medium or Large. Medium and Large differ in compute, and Medium is the faster choice for iteration. Local routes give full control and a portable .safetensors file. General LoRA basics are covered on the LoRA page.

What settings work locally?

The community Krea2Trainer wrapper drives ai-toolkit through a web page. It has an MIT license and dates from June 2026. Its defaults are 2000 steps and a learning rate of 0.0001. Rank and alpha are both 32. Batch size is 1. Eight-bit quantization fits 24 GB cards. The base download is about 50 GB as a diffusion pipeline folder.

Trigger words and per-image .txt captions are expected, as in other training diagnostics setups.

What goes wrong, and what is still unclear?

Open ai-toolkit reports list a gradient-checkpointing crash (issue 1041). Another report describes a 12 GB card running slowly (issue 1032).

Preview samples cannot yet use the Turbo weights while training on Raw. The wrapper README names an older base architecture. The ai-toolkit source instead shows a 28-layer diffusion transformer, so the wrapper settings are community defaults, not official ones. Finished files load through ComfyUI loader nodes.

Sources