Krea launches LoRA training for Krea 2
Krea has opened beta LoRA fine-tuning for its in-house Krea 2 image model, so paying users can train their own styles, characters and objects and reuse them in its Image tool.
Krea announced that LoRA training for its Krea 2 image model is now rolling out in beta. For now it is limited to subscribers on the Max and Business plans.
Krea already supported LoRAs on third-party backbones, including Flux, Qwen and Wan. The company says Krea 2 LoRAs hold "style, character, and material identity" better than any of those earlier backbones did. It offered no benchmarks or side-by-side results to back that up.
Users need at least three images that share a style, character or object. They then choose a base model: Krea 2 Large, which Krea says trains stronger, or Krea 2 Medium, which trains faster. The step count is adjustable, and the default is tuned for most datasets. Before training, Krea captions the dataset automatically so one-off backgrounds or props are not learned. Users can edit those captions first, and Krea shows an estimated compute cost before a run is submitted.
Finished LoRAs show up in a Train library with preview images. They can be applied in the Image tool when Krea 2 is selected, with adjustable strength, and several can be stacked, such as a character LoRA combined with a style LoRA.
Trained LoRAs are private by default. Owners can share them with collaborators or make them public, which lets other Krea users generate with them. Krea says the original training images stay private either way.
Krea did not say when the beta will end, whether the feature will reach other plans, or how much training costs beyond the compute estimate shown in the product.