Dit layers in training
Hello everyone, I have a question: when starting training in the Anima standalone trainer on a custom Checkpoint other than the Base version, I get an error about missing DIT layers. I thought this was only a problem with unofficial checkpoints, but when I decided to test it on the new "Aesthetics and Turbo" versions, I got the exact same error. I'd like to ask the experts if this is a feature or just a current issue with this version?
If you're using sd-scripts, this pull request should resolve the issue. https://github.com/kohya-ss/sd-scripts/pull/2404
I can't say for sure, because I've always used the Base model to train Lora, and that's what the official documentation recommends as well.
If you're using sd-scripts, this pull request should resolve the issue. https://github.com/kohya-ss/sd-scripts/pull/2404
I can't say for sure, because I've always used the Base model to train Lora, and that's what the official documentation recommends as well.
I tried different training methods, both Anima Base and Anima Aesthetic, and also tried different types of merges and finetune, like Nova or Janima. They all crash with roughly the same error, and the assistants all responded that it was a problem with the lack of dit layer's. I also tried different training tools "Kohya_ss, Anima Standalone Trainer", but the result was the same. That’s why I became interested in where the error occurs, in the lack of adaptation of training or The problem lies within the architecture itself