What is your training secret?
Dude, this model is amazing! I tried to merge it with NAI with a 30/70 ratio and have gotten some pretty amazing result. Mad respect to Yoshitaka Amano and you! Would you please share your train secrets, like how many sample image you used and what checkpoint did you use?
Dude, this model is amazing! I tried to merge it with NAI with a 30/70 ratio and have gotten some pretty amazing result. Mad respect to Yoshitaka Amano and you! Would you please share your train secrets, like how many sample image you used and what checkpoint did you use?
Thanks,
From my limited testing it really come down to the Regularization Images
Instead of using person_ddim Regularization Images, I generated 1500 images from a prompt that produce image close to the desired result.
ex: final fantasy landscape, colorful, black background, vibrant colors, high contrast, many thin pencil strokes contour, watercolor"
I try to alternate few keywords (ex: Final fantasy landscape or woman or on white background or knight) and produce 128 images each time.
Then I select 20 images from the real artist to put in the training images folder and start the training.
I will try your merge to see the result
I tried you mix and I like the outcome. You can produce great 80s style cartoon with it.
Thank you for sharing! It is interesting how the mix doesn't really look like his work at all, but the quality improves quite a bit even without any words that trigger his art style.