OneTrainer - how to use with float16

First: very, very thanks to

Stable Cascade Lora Training with OneTrainer | Civitai

for tutorial. It was hard to understood, but finally it working :)

Dont worry! It working on RTX 3060ti 8gb vram. I have old ryzen 7 with 32gb ram. (training loras and run workflow)

Working lora:

Stable Cascade Beksinski - v2.0 | Stable Diffusion LoRA | Civitai

Presets and workflow added as zip file.

So, lets go!

Put json (presets file) setting to folder

Step by step:

I have images and txt files on drive L. bfloat32 is so fun, but so slowly on my gpu, so i switch all to f16 (not bf16)

I was use dataset tool to make captions with BLIP and set trigger word as first in caption.

Caption tool tip: first use Blip to make first line with tag and caption, then make wd14 caption as NEW LINE. Results:

"beksinskiart, a painting of a man kneeling down in the sand

traditional media, fire, no humans, solo, border"

If you need longer caption you have to add max_tokens in blip or blip2 model.

and last one window

Testing workflow:

And normal workflow

Have a nice day!

After training then you use my wf with this prompt

"beksinskiart, a painting of a man kneeling down in the sand

traditional media, fire, no humans, solo, border"

you got this: