Home Artists Posts Import Register

Content

This Is How I Am Doing Research To Find Best Training Hyper Parameters For SDXL DreamBooth Training

Patreon exclusive posts index

Join discord and tell me your discord username to get a special rank : SECourses Discord

Started 6 different Automatic1111 SD Web UI on a 6x RTX 4090 GPUs having RunPod machine.

Doing x/y/z checkpoint comparison to find best hyper parameters for Stable Diffusion XL (SDXL) DreamBooth training with:

U-NET only
vs
U-NET + Text Encoder 1
vs
U-NET + Text Encoder 2
vs
U-NET + Text Encoder 1 + Text Encoder 2

Comparing lots of checkpoints with lots of different prompts.

Being sure to not overtraining and cooking model with Text Encoder training is super important.

Hopefully the results will become a public tutorial video later.

The new found config will be posted here : https://www.patreon.com/posts/very-best-for-of-89213064


Files

Comments

mypatreonemailacc

Is there a reason "Enable Buckets" is disabled in the 24GB config?

Furkan Gözükara

i prefer to train without them but of course you can enable it if you have different aspect ratios. the effect being different

Anonymous

When you are evaluating a lot of different checkpoints and prompts do you still use auto1111 or something more automatic to run the whole xyz plot? I tried to come up with a diffuser workflow but I don't get the same results as in auto1111

Furkan Gözükara

I am using Automatic1111. Diffusers is really really harder to replicate to get same results as Automatic1111. Auto1111 has so many additional improvements

mypatreonemailacc

Have you tried any of the "no parameter" optimizers like Prodigy or DAdapt? Do you recommend them?

mypatreonemailacc

Is it possible to start the training from a fp16 checkpoint from civitai? Or do we have to use the SDXL base model?

Furkan Gözükara

it depends on purpose. for example for stylized output some civitiai models are better. i say compare both sdxl base and civitai model