Krea opens LoRA training for its Krea 2 model
By Steven Van ·
Max and Business subscribers can now train custom styles, characters, or objects and apply them in the Image tool.
Krea has opened LoRA training for its Krea 2 model, in beta for Max and Business subscribers. It lets you train the model on your own images to capture a style, character, or object, then use it directly in the Image tool.
Training happens at /train: upload at least three images sharing the style, character, or object you want to capture, choose between Krea 2 Medium (faster) or Krea 2 Large (stronger), and adjust the step count if needed. Krea auto-captions the dataset before training, and captions can be edited beforehand. Once training finishes, the LoRA appears in the Train library and can be applied from the Image tool with adjustable strength, and multiple LoRAs can be combined, such as stacking a character with a style.
LoRA training already existed for Flux, Qwen, and Wan, but Krea says Krea 2 LoRAs hold style, character, and material identity more precisely than the earlier backbones. Trained LoRAs are private by default; a style can later be made public for others to generate with, though the original training images stay private.
Read the full announcement or try it on the Krea product page.
