Technidrawing
Dreambooth based on ~730 images of specific style.
Forces very deep blacks and bright whites in image, as it is a feature of style.
Usage
It often works really good with latent upscaler (most of preview images are upscaled with latent), so, maybe you want to give it a shot.
Otherwise - same as with other models, 2m karras 20 steps will get you a long way.
Training
Similar to previous trained checkpoint(no shit, they are both styles based on almost same dataset).
Dataset
~730 images, owned by me.
Description
FAQ
Comments (3)
The unevenness like leather fabric and the horizontal lines like investigation lines stand out, so it is difficult to handle unless you take advantage of them or kill them. The author's sample gives the impression of killing it. The contrast between black and white and black and other colors is strong, so it is suitable for this purpose. However, monochrome images encounter the aforementioned peculiarities.
Rather than being able to create beautiful images, it would be better to think about how to use a model that has its own peculiarities.
It is very peculiar. For example, [evil grin] was interpreted as [grin] by [evil], and it is difficult to get the effect of LORA, which erodes the picture pattern.
I also reused images that could not be uploaded to the bounty (I was unaware of the specification that did not allow them to be added).
クセ強め。例えば[evil grin]を[evil]が[grin]すると解釈した(悪魔っ娘になった)り、絵柄を侵食するLORAの効果を受けにくいなど。
あとバウンティの方にアップできなかった(追加できない仕様を知らなかった)画像を再利用。
I really liked this contrasting unsaturated graphic hand-drawn style! A very expressive, stylish and at the same time flexible model.
Great job, Thanks!