8 step issue with other loras
#22
by Plaguekind - opened
First of all 8 step is amazing, good job.
However if you use it with other loras, it breaks the distillation of the model.
Is there a fix/workaround for using other LORAs with turbo??
AFAIK LoRa is addition to the base model, so using other LoRa(s) that wasn't trained on distilled model will bring back the non-distilled weights, thus degrading the distillation from Turbo LoRa.
You can try increasing the strength of the Turbo LoRa to get stronger distillation (aka. allow less steps) against those non-distilled LoRa(s), but those non-distilled LoRa may not be perfect/accurate anymore (similar to using multiple Character LoRas that affects each other accuracy).