Training on new concept
Hi, Thankyou for making this opensource.
I have a query on how to train or fine tune it for new concept. Do I need to train base models of SD1.5 or SDLX or I have to train the unet or LORA.
Is training also going to be turbo fast?
Sorry if it sounds stupid, as I am new to this.
another query is can we pass two images of different people and get combined image?
Hi,
@gulshansaini
For fine-tuning, please refer to https://huggingface.co/ByteDance/Hyper-SD/discussions/35#66441d450c34964ea13b125c.
It should be also fast at inference after fine-tuning if utilizing our models.
Hi Thankyou for response. I have already gone through the mentioned thread yesterday however it is not clear to how to proceed. Do you have some working examples that I can look into?
You have mentioned two ways
- Try merge it into UNet and create a new LoRA
- just fine-tune our LoRA
I would appreciate if you could help in anyway
Hi,
@gulshansaini
These are implementation issues. You could refer to diffusers api to see how to make it happen.
Our LoRA is in kohya format and you would need to convert it into peft format and load for training on GPU.