DarkViewAI

WARNING gradient_accumulation_steps is 3. accelerate does not support train_db.py:109

1

WARNING gradient_accumulation_steps is 3. accelerate does not support train_db.py:109 gradient_accumulation_steps when training multiple models (U-Net and Text Encoder), so something might be wrong WARNING gradient_accumulation_stepsが3に設定されています。accelerateは複数モデル（U train_db.py:112 -NetおよびText Encoder）の学習時にgradient_accumulation_stepsをサポートしていないため結果 は未知数です

Extracting Lycoris For SDXL Issue with Text Encoder 2

2

text encoder 1: _IncompatibleKeys(missing_keys=[], unexpected_keys=['text_model.embeddings.position_ids']) text encoder 2: _IncompatibleKeys(missing_keys=[], unexpected_keys=['text_model.embeddings.position_ids']) building VAE loading VAE from checkpoint VAE: building U-Net loading U-Net from checkpoint U-Net: building text encoders loading text encoders...

Full FP16 Dreambooth for SDXL does not work

2

Either produces nans, or doesn't train at all, all the sample images are the same.

Im not sure if its only for me, but with the latest kohya it seems the dreambooth extraction is providing worse results. Extracting with an earlier version 22.6.2 seems to...

any way to disable scheduled huber loss

5

is there any way to remove it or disable it from using?

does sdxl lora training train both text encoders or only 1?

3

results with dreambooth and text encoders are pretty unstable and unreliable. but when i train via lora and with the text encoder results are pretty good. so i'm wondering if...

Is it possible to do this? For example I have 20 - 1024x1024 Images of someone I have 10 - 832x1216 Images of someone I have 5 - 768x1344 Images...

any plans on adding it?

Does sdxl lora text encoder training, train both text encoders? or just one?

Loras are getting patched after each generation

1

it patches it after each generation