Generated images

[Ishihara-Masabumi]

I tried to train and generate images as the following command lines.

!python imagen.py --train --source datasets --imagen model6.pth --train_unet 1 --no_sample --no_elu --epochs=80
!python imagen.py --train --source datasets --imagen model6.pth --train_unet 2 --no_sample --no_elu --start_epoch=81 --epochs=160
!python imagen.py --imagen model6.pth --sample_unet=2 --size 256 --output girl_256.png --tags "1girl, blonde_hair" --cond_scale=1.1 --replace

The first generated image is below.

The second generated image is below.

Are they OK? or lack of training?

[deepglugs]

What does --sample_unet=1 look like? Depending on which unet looks better, the other will need more training.

[deepglugs]

Here's some of my samples from my new model after 40 epochs with 107k images:

unet1, 256 dim. cond_scale=1.1

Even after 40 epochs (4.3 million images seen) I need a lot more training.

[RahulBhalley]

@deepglugs How long it took you to train this much and what GPU are you using?

[Ishihara-Masabumi]

The training time is about 2 or 3 days wih NVIDIA A10.

[RahulBhalley]

Woah! In paper they trained for 2.5M steps with a batch size of 2048 (are u guys using same batch size?).

[RahulBhalley]

What dateset are you using? The LAION-400M seems 10 TB. That's HUGE.