DarkViewAI

Can't fit on a 24gb gpu, but kohya sd scripts says it works with 22-23gb vram

1. I am getting worse results with t5 attention mask when using flux shift as my timestep. Better results without Apply T5 Attention Mask. 2. I spoke to kohya and...

multires training with toml seems to learn nothing

1

tried training my dataset in 3 different resolutions, trained for 5k steps, it learnt nothing.

Flux sampling taking 4-5 mins on wandb with 1024 training

1

when training with 1024 on flux while also having wandb turned on, it seems my sampling is taking 4-5 mins to sample one image, not sure why.

is it possible to implement an easier system that we can use, that automatically takes one dataset and buckets it into 512,768, and 1024

Flux Dreambooth crashes after trying to save checkpoint

1

I am able to successful db training and samples look good, but after 600 steps it gives me this error 2024-08-19 07:57:06 INFO train_util.py:5123 INFO saving checkpoint: train_util.py:5124 /home/Ubuntu/Downloads/model 4/Rick-step00000600.safetensors...

Flux dataset training higher than 1024 pixel images = bad results

4

When my images in my dataset are higher than 1024, like 2048 or 2536 the likeness in my training is gone. is there any way to fix this or is...

Is it possible to implement these features from ai toolkit into kohya. Fully explained down.

6

I have done over 300+ trainings now (renting gpus) of different training parameters from both kohya and ai toolkit. I get good to decent results now on both. I believe...