FedorAiToolkit has been introduced—a specialized fork of ai-toolkit that uses the DRaFT-K method to train LoRA and LoKr models based on Krea 2. The tool achieves unprecedented accuracy in character creation, capturing not only facial features but also unique anatomical body proportions.
What Happened
Developers have introduced FedorAiToolkit, which implements a two-stage training process. Following a standard SFT (Supervised Fine-Tuning) stage, an optimization stage follows using differentiable rewards for facial similarity via ArcFace and for body geometry via SAM 3D Body. On an RTX 5090-level GPU, a full training cycle takes approximately 60 minutes.
Context
Traditional LoRA training methods often focus on textural and pixel similarity, which leads to a loss of anatomical consistency when generating full-body characters. The DRaFT-K method shifts the focus to semantic optimization, using high-level features (identity and geometry) as target metrics.
Why It Matters for the Industry
The implementation of differentiable rewards allows the industry to move from simple pixel imitation to optimization based on semantic features. This paves the way for standardizing the creation of highly consistent digital humans and integrating such methods into cloud APIs for automated control of geometry and identity.
Why It Matters for Users
For content creators and researchers, this means the ability to generate stable characters that maintain their physique and proportions across all angles. The tool makes deep personalization accessible to individual authors, providing professional-level quality within the Krea 2 ecosystem.
What Is Not Yet Known / Limitations
The current implementation may not be sufficiently applicable in corporate environments due to a lack of built-in control and data management mechanisms, making it geared more toward individual researchers and solo creators.
Sources
Author
Look at AI, Editorial Team