LoRA Training Beta Launches for Krea 2 Models

Krea, the AI image generation and enhancement platform, launched a closed beta for LoRA training compatible with its Krea 2 Medium and Krea 2 Large models. This update integrates the Krea Train tool directly with the latest model architecture, allowing users to define specific visual styles, characters, or products that the AI can then replicate across multiple generations. By training these Low-Rank Adaptation (LoRA) weights, creators can move beyond generic prompting to achieve precise aesthetic control.

What's new

Krea 2 LoRA training is currently rolling out as a closed beta within the Krea Train interface as of late 2024. During this initial phase, selected creators and teams can train custom weights for both the Krea 2 Medium and Krea 2 Large models. The update introduces a toggle for privacy, giving users the choice to keep their trained styles restricted to their private workspace or publish them as Public LoRAs for the wider community to discover and use within the Krea Image generation tool.

This training capability focuses on five primary categories: brand styles, specific characters, product photography, material studies, and personal subjects. Unlike previous iterations that relied on more generalized models, these LoRAs are optimized specifically for the Krea 2 architecture, which aims for higher fidelity and better prompt adherence. The rollout allows for a more integrated workflow where a trained asset can be immediately applied to a generation without switching between disparate platforms.

How it fits your workflow

Krea 2 LoRA training provides a solution for the consistency issues that often plague AI video and image production. For art directors and concept artists, this tool replaces the need for complex prompt engineering to maintain a character's appearance across different scenes. By training a LoRA on a specific character design, a filmmaker can ensure that the protagonist looks identical in every concept frame, a workflow similar to using custom GPTs or fine-tuned Stable Diffusion models.

In a professional production environment, Krea 2 serves as a viable alternative to Flux.1 or Midjourney for teams that require bespoke style consistency. While Midjourney offers style references, Krea's LoRA approach allows for a deeper, more permanent technical 'memory' of a visual asset. This is particularly useful for storyboard artists who need to generate dozens of frames in a unified brand style. The ability to share Public LoRAs also mirrors the ecosystem of Civitai, potentially turning Krea into a hub for specialized architectural or cinematic styles that other editors can utilize for rapid prototyping.

What it costs / how to try it

Krea 2 LoRA training is currently restricted to a closed beta for selected users and teams. While the standard Krea Image and Enhance tools are available via various subscription tiers, access to the Krea Train beta for Krea 2 models requires an application or an existing professional-tier account during the testing period.