Release of FedorAiToolkit for Precise LoRA/LoKr Training

FedorAiToolkit has been introduced—a fork of ai-toolkit that utilizes the DRaFT-K method for high-precision optimization of Krea 2 models based on facial features and body geometry.

Compiled by Sergey KostenchukPublished 2026-07-22Updated 2026-07-22

2026-07-22 Multimodal

🎨 FedorAiToolkit has been released—a fork of ai-toolkit implementing the DRaFT-K method for training LoRA/LoKr on the Krea 2 model.

The tool uses differentiable rewards (ArcFace and SAM 3D Body) to optimize body geometry and facial similarity. This enables high-precision reproduction of full-body characters. On an RTX 5090, a cycle takes ~60 min.

🌍 The transition to semantic feature-based optimization improves the quality of model personalization.

👤 Creating consistent characters with unique physiques becomes easier.

Source 1: https://github.com/CliffNodes/FedorAiToolkit Source 2: https://github.com/FugueSegue/bodyrec

Sources