TL;DR:
Kohya is in Training Studio. Set the Engine to Kohya to train SD 1.5, SDXL (including Pony, Illustrious and NoobAI), Flux.1 and Chroma LoRAs with it.
AI Toolkit stays the default. Kohya is one tap away, next to the media and type you pick at the start.
The settings you know are there. Optimizers, schedulers, repeats, bucketing, clip skip, network dim and alpha, and Train Further and publishing work the same as any other run.
The classic trainer is retired. It stopped starting new runs on November 16th, and as of today there's no switching back to it.
Your published models aren't touched. They stay up exactly as they are.
Here it is in 35 seconds. (And yes, this video was made by AI)
When we moved training over to Training Studio, Kohya didn't come with it, and plenty of you let us know. It had already stopped working in the classic trainer in October, so we switched it off there rather than let your runs fail. That left a lot of people who trained with Kohya for years without the engine they trusted, and I get why that was frustrating.
Why it's in Training Studio
We didn't want to patch Kohya back into the old trainer. We wanted it in Training Studio, where you can watch a run as it goes, compare every epoch side by side, and generate with any of them before you decide which one to keep. So the work was making Kohya a real part of the Studio, with its own settings, its own pricing and its own loss graph, rather than a checkbox bolted on.
How it works
Start a new training the way you normally would. Next to the media and type at the top there's an Engine choice, AI Toolkit or Kohya. Pick Kohya and the model list narrows to the ones it can train, which today are SD 1.5, SDXL and the SDXL families (Pony, Illustrious, NoobAI), Flux.1 and Chroma.
Everything after that is the same flow as any other run. Bring in your images, let them auto-label or write your own, and set your steps. On the Review step you'll find the Kohya settings people kept asking about, including AdamW, AdamW8bit, Adafactor, Prodigy and Lion optimizers, the usual schedulers, repeats, bucketing, clip skip on SD 1.5, and fp8 and block swapping on Flux.1 and Chroma. You see the price before you start, and once it's running you can follow the logs and the loss graph and generate with each epoch as it lands.
When it's done, Train Further and publishing work just like they do for an AI Toolkit run.
The classic trainer is retired
The classic trainer stopped starting new runs on November 16th. Until now you could still flip back to it from your settings and poke around, and that switch is gone now. Training a LoRA always takes you to Training Studio.
Nothing you trained there is lost. Your classic runs and drafts are still on the Training tab of your profile, and you can still open them, review them and publish them. Every model you published from it stays up exactly as it is. Training Studio's own list shows your runs from the last 30 days, so publish or download anything you want to keep.
Tell us how it trains
If you trained with Kohya before, I'd love to hear how your first Studio run compares, good or bad. Drop it in the comments, especially if a setting you relied on is missing, since that's the quickest way for us to know what to add next. Thanks for sticking with us through the switch.





Comments