Some quick experiments with the new open source v1-alpha AI model, upscaled with ESRGAN. Which is basically an open source version of

Takes around 4s per image and 10GB VRAM on RTX3090. Great to see a new open source model.

Not bad! And no it doesn't do hands!

0 7

Testing updated CLIP Guided Diffusion code. 512x512 pixel results upscaled x4 with Real-ESRGAN.
"a pastel of a demon"
"an ugly man"
"an ultrafine detailed painting of Al Pacino"
"Brad Pitt made of mist"

1 16

Some experimentation with 6 graphics and It sometimes requires a little manual aliasing on some solid lines, but the result is quite good!

2 25

training a 'sequel' to manga109attempt for ESRGAN.

JPG/dither friendly, trained on photos and traditional/digital art, 2x upscale (only nn/bicubic filtering should go beyond 2x imo), looks a bit soft but 'safe', doesn't radically alter the image

18 hours, 0 epochs ._.

0 2