Krea 2 Sees a Wave of Lightweight Style LoRA Experiments

From July 16 to 17, multiple community posts focused on LoRA training with Krea 2. The significance is not an official model update, but the fact that users are already treating Krea 2 as a base model for frequent, lightweight style fine-tuning, with experiments spanning still images, illustration styles, and video-like outputs.

Representative cases

@Jolly-Rip5973 shared a Krea 2 LoRA trained on the style of 1920s illustrator Ida Rentoul Outhwaite. According to the author, it can produce black-and-white pen/ink illustrations, watercolor images, and smoother children's-book-like results. @Beautiful_Egg6188 said they have recently been using Krea 2 almost as often as Flux1.Dev, and after seeing an Instagram short video, trained a LoRA for a specific visual style and shared the model link. @XIII-TheBlackCat posted a gallery of Krea 2 LoRA test images, but did not provide additional methodological details.

Training setup and output range

@car_lower_x provided the clearest training details: they used Ai-Toolkit on KREA2 Raw to train a Supermarionation LoRA with only 40 low-resolution still frames. On the following day, the same author showed video-style results from a KREA2 LoRA trained on stills from 1960s and 1970s Supermarionation TV shows. In another post, @freshstart2027 shared locally generated Krea 2 animal illustrations made with a personal LoRA stacked on top, mainly as workflow and visual-reference examples.

Taken together, these posts do not yet form a shared tutorial or standard parameter recipe, but they do show that Krea 2 has quickly entered community LoRA experimentation, that training data can be quite lightweight, and that the output is not limited to single still images.

2026-07-16 ~ 2026-07-17 · 6 related posts