Read one day of our news radar and you would not see this. Read two months and it is there in the article titles alone. Kohya is down on 23 August. Future LoRA training and the Kohya problem in September. Is Kohya fixed? on 2 October, Kohya is now totally dead on the 16th, and this morning, Kohya is gone and what’s next.

All of those are posts by Civitai creators, and all of them are about Civitai.

What happened, from Civitai’s own code

Civitai’s site code is public, and a pull request merged on 21 October spells it out. When a feature flag called kohyaTraining is off for a user, any training run set to the Kohya engine is now refused before anything is charged, with a message asking the user to switch it to AI Toolkit. Flux training defaults to AI Toolkit. In the pull request’s own words, the flag “remains the switch for retiring Kohya.”

The same description explains a bug creators had been reporting for weeks. Before the fix, a run saved earlier with Kohya selected could still be submitted, and no worker would ever pick it up. One creator wrote on 14 October that Kohya jobs sat at “Processing” and failed only after 24 hours. That was the gap the pull request closed.

Before the code there was a cap. On the 18th a creator reported that the on-site Kohya option had been limited to 20 steps with no way to resume, which ended training for characters with complicated outfits. This morning another reported that the option to select it is gone.

We found no announcement. Creators have said as much themselves, and one asked for more transparent communication from the people running the site. As far as we can tell, the code change is the clearest statement Civitai has made.

What did not happen

The open-source Kohya trainer, the scripts the whole on-site option was built on, had a commit this morning. It has more than 7,000 stars and is not archived.

AI Toolkit, its replacement on Civitai, had a commit yesterday.

The one soft spot in the local Kohya stack is the popular graphical front end for those scripts, which has not had a commit since 1 August. If you run Kohya through that interface, that is the thing to watch, not the headlines.

What creators actually lost

Read the posts closely and nobody is mourning a piece of software. They are mourning a button.

The creator who wrote this morning says it directly: training on their own PC is extremely labour intensive, and the on-site trainer is why they used Civitai at all. Another tried to train the same LoRAs locally with the settings that worked on the site and could not get comparable results. Several report that their Illustrious character LoRAs come out worse on AI Toolkit. That is their verdict, not ours. We have not compared the two.

So the loss is two things. The convenience of somebody else running the trainer. And a set of working settings that lived inside somebody else’s form, with no guarantee they transfer anywhere else. AI Toolkit does not take Kohya settings as a drop-in. The pull request itself notes that it seeds AI Toolkit with its own step-based settings.

If you trained this way

Write your settings down while you can still see them: network size, steps or epochs, learning rate, base model, captioning. A configuration you can only see in a hosted form is one you can lose without notice.

Keep your datasets yourself, outside the site. The images and captions are the part you cannot rebuild.

If Kohya specifically is what works for you, it still runs locally, on the same scripts the site used. The setup is real work. It is also the only version of Kohya that cannot be switched off by a feature flag.

We train our own LoRAs locally with a different trainer, so none of this touches our pipeline. It is the same lesson we drew from a node pack that vanished last month, arriving from the other direction: the tool was fine all along, and the thing that disappeared was somebody else’s convenience wrapped around it.

A daily note from our news radar. The week gets the full treatment in the DIY AI Brief every Monday.