Civitai has officially decided to drop long-term support for LoRA training using the Kohya engine. As of September 17, 2026, training with Kohya is technically still available, but in practice, it barely works anymore. Instead of Kohya, Civitai is heavily pushing AI Toolkit. However, looking at the training results, I have personally never felt that AI Toolkit holds a candle to Kohya—at least when it comes to SDXL-based models, even if it performs decently on other baselines.

For this reason, I’m putting my LoRA production on hold for the time being (and may even consider retiring). While AI Toolkit handles certain other models fairly well, I haven't found a single base model that truly hits the mark for me yet. So, I’ve decided to wait until a proper, worthy model comes along.

When that time comes, I’ll be back. Until then, take care and goodbye...