Custom LoRA Training Arrives for Krea 2 Beta

Krea has introduced LoRA (Low-Rank Adaptation) training for its Krea 2 model, currently available in beta. This update allows filmmakers and digital artists to move beyond generic prompting by training the AI on specific visual datasets. By providing a set of reference images, users can now bake a particular character, object, or art style directly into the generator's architecture for consistent output across multiple frames or projects.

What's new

The primary addition is a dedicated training interface within the Krea platform. Previously, users relied on style references and moodboards to influence the look of an image, which often led to variations in detail. With LoRA training, the system learns the underlying geometry and texture of the provided samples.

Key capabilities include:

• Character consistency: Train the model on a specific person or creature to maintain their likeness across different environments.
• Object permanence: Teach the model to recognize a unique prop or vehicle for use in storyboarding.
• Style locking: Define a specific cinematic color grade or illustrative style that remains uniform regardless of the prompt.
• Direct integration: Once trained, the LoRA can be toggled and adjusted within the standard Krea 2 image generation tool.

How it fits your workflow

For filmmakers and concept artists, this feature addresses the problem of "AI drift," where characters or styles change slightly between generations. In a pre-production workflow, a director can train a LoRA on a lead actor's headshots or a specific costume design. This allows the storyboard artist to generate hundreds of frames where the protagonist looks identical in every shot, a task that was previously difficult with standard AI video generation and image tools.

Editors and VFX artists can use this to create consistent background plates or matte paintings that match the lighting and texture of a specific film stock. While tools like Civitai or local Stable Diffusion setups have offered LoRA training for some time, Krea simplifies the process by hosting the compute and providing a streamlined UI. It serves as a middle ground between the high technical barrier of local training and the limited control of basic text-to-image platforms like Midjourney.

What it costs / how to try it

LoRA training is currently rolling out in beta to Krea users. Access typically depends on your subscription tier, as training requires significant server resources. You can check your account dashboard on the Krea website to see if the training module is active for your profile.