Train a subject face in the Dreambooth extension using any SD v2.1

[TypicalPlay]

Has anyone trained a subject face in the Dreambooth extension using any SD v2.1?

Resolution 768

This on a 3060 12gb

NOTE - NOTHING TO DO WITH LORA

[poondoggle]

Not without unchecking text encoding and that renders 2.1 useless from the tests I've done with my 3060.

[TypicalPlay]

Yes, was able to complete the training for SD 2.1 with the following setting's on a 3060 12gb

Memory Attention - xformers
Train Text Encoder - UNCHECKED

The results were a bad resemblance of the subject's face.
But on one occasion setting the CFG to 5 it came out brilliantly.
BUT, you cannot go about pressing the Generate button a million times to get these rare images.

---

Today was able to complete the training for SD 2.1 with the following setting's on a 3060 12gb

Memory Attention - flash_attention
Train Text Encoder - CHECKED

ALL THE SAMPLE generated as part of training were black.

Stopped midway. It was taking a lot of time.

BUT no cuda memory errors

[zahdab]

for black images, try setting memory attention to xformers when training ... it worked for me... i trained for 768 model using ... though the model lost a lot of detailed ... even if i dont promp the subject everything became unrefined ... l