Use with the Perturbed Attention Guidance node for best results. I use a value of 2.6
V26 is a complete re-train and an incremental update over v19. It should be a little bit more stable at the default strength of 1.0, but will still produce a lot of nightmare fuel. You might prefer v19 for more flexibility and diverse outputs.
It was trained for 4600 steps with multi-res and then finished off for 2000 more steps at 1024. This is also a merge since even after the 6600 total steps the effect was weak.
Despite cropping, it also tends to introduce false watermarks.
Use with the Perturbed Attention Guidance node for best results. I use a value of 2.6
V26 is a complete re-train and an incremental update over v19. It should be a little bit more stable at the default strength of 1.0, but will still produce a lot of nightmare fuel. You might prefer v19 for more flexibility and diverse outputs.
It was trained for 4600 steps with multi-res and then finished off for 2000 more steps at 1024. This is also a merge since even after the 6600 total steps the effect was weak.
Despite cropping, it also tends to introduce false watermarks.