Requesting additional training methods.
LECO: "erase concept" training, it's low priority in my mind but also sort of helpful as a way of merging minor refinements into existing LoRA
ADDifT: Difference training method that allows you to use a very small dataset (as low as two images) to capture a concept with little to no style/composition bleed. It also has less training steps required compared to "copier" difference training methods which involved frying a lora and subtracting.
Here is an extension that has some visual examples and working code for these training types
I don't expect LECO to be implemented because IIRC it came up before in the main branch as being an issue and I feel like it has a much longer train time compared to other methods. But either way I do appreciate you taking the time to even read this request. 🌸
Requesting additional training methods.
LECO: "erase concept" training, it's low priority in my mind but also sort of helpful as a way of merging minor refinements into existing LoRA
ADDifT: Difference training method that allows you to use a very small dataset (as low as two images) to capture a concept with little to no style/composition bleed. It also has less training steps required compared to "copier" difference training methods which involved frying a lora and subtracting.
Here is an extension that has some visual examples and working code for these training types
I don't expect LECO to be implemented because IIRC it came up before in the main branch as being an issue and I feel like it has a much longer train time compared to other methods. But either way I do appreciate you taking the time to even read this request. 🌸