Player not loading? Watch on YouTube
In this November 2025 Russian-language tutorial, the speaker covers Flux face LoRA training in Kohya_ss, installed through Stability Matrix on Windows 11. The speaker chooses a portable installation and recommends an SSD location with no spaces, Cyrillic characters or special symbols in the installation path. Stability Matrix also gets a brief overview as a manager for installing interfaces and organizing models and generated images.
The setup uses an FP8 Flux model alongside AE, CLIP-L and T5 model files. The speaker explains how consistent filenames keep preset paths reusable, then separates each training project into folders for images, logs, outputs and optional regularization images. Captioning produces .txt files with a shared trigger prefix. The supplied preset and additional lessons are available through the creator's Boosty page.
The example has 20 images with two repeats each and 50 epochs, targeting 2,000 steps. The speaker recommends clear faces with varied clothing and backgrounds, and compares 512 × 512 test training with 1024 × 1024 for facial detail. Settings include Adafactor, latent caching, fixed seeds and checkpoints with test samples every 10 epochs.
Troubleshooting focuses on incorrect model paths and cache conflicts with augmentation or caption changes. Resuming training requires a saved state. The speaker judges the example's 2,000-step result favorably, while noting that other faces may need different settings.
