Train a full checkpoint - not LoRa

[Samael-1976]

Hi, I'm really curious about pixart.

I've just fee question:
Premiss: I will use OneTrainer

1. There is a tutorial for the training?
2. I'm with 2060 with 12gb of vram, I will able to train a 1024 model (just for saying that I found a way to train sdxl with my card).
3. Any other suggestion to give me?

thank you for any answer!

Samuele

PS: sorry for my horrible english

[chrish-slingshot]

Hey. Your best bet for training is to use OneTrainer - they just added PixArt Sigma.

https://github.com/Nerogar/OneTrainer

I don't think there's a tutorial for the training but if you hit me up on Discord I can give you some pointers - username _.crash._. There's also the PixArt discord:

https://discord.gg/MfZFVKfCwD

12GB should be enough. You won't be able to train the text encoder, but then no reasonable commercial system can at the moment. I'm running a 1024 fine-tune at the moment with a batch size of 16 and hitting 17GB, so with a lower batch size (i.e. longer training time) you should be fine.

[Samael-1976]

A thousand thanks!
I just sent you a friend request on discord (I'm Samael1976).

Any help you can give me will be well accepted, since I know very little about pixart. But since it came out I've only read great things and it's always intrigued me.

I already use OneTrainer for XL training :)