Training LoRAs Without the Command-Line Hangover

What it does This is a PySide6 desktop wrapper around Kohya’s sd-scripts. It exposes the backend’s LoRA, LoCon, and LyCORIS training options through collapsible panels, drag-and-drop file selectors, and a TOML-based config system so you don’t have to assemble long command strings by hand.

The interesting bit The queue system treats training like a playlist: you can stack jobs, reorder them with arrow keys, and edit args or add new items while another job is already running. The author also added a custom CosineAnnealingWarmupRestarts scheduler that spreads warmup steps across epochs and decays the learning rate after each restart.

Key highlights

• Block-weight training, subset management, and SDXL support exposed through the UI.
• TOML save/load for configurations; drag-and-drop or dialog-based file selection for most paths.
• Live queue manipulation during active training runs.
• Theme switching via qt-material that persists across restarts, plus a “no-theme” fallback for checkbox rendering issues.

Caveats

• Advanced optimizer arguments still require hand-editing the TOML file; the UI for them is noted as coming later.
• Configs from the old script generation are incompatible; previous JSON save files will not load.
• The application is fundamentally a frontend and installer glue around Kohya’s sd-scripts; it will not train without the backend.

Verdict Worth a look if you train LoRAs regularly and prefer a desktop queue to terminal fatigue. Look elsewhere if you need a standalone training engine rather than a GUI wrapper for Kohya’s scripts.