Help push style by using negatives: photo, blurry, render
What's different?
Smaller dataset with flip augmentation
More responsive to prompts
Stiff in variety of outputs
Smaller file size for model
Trimmed down the dataset and cleaned up art via vector conversion. Edges are crisper, less colors, more separated. This version is kind of a side grade. It isn't exactly better than the previous or worse as a whole. Parts are better, like this version is more responsive, but it's also harder to push the style. The dataset has less variety so the model is a bit stiff. It can sometimes be impossible to get the style to push properly even with correct triggering in prompt. IMG2IMG is best for pushing the style on subjects that it just can't generate with prompting alone.
The sampler and scheduler can be really tricky with this model. With some scheduler sampler setups it will look like the model isn't doing anything at all or is just making really bad images, blurry mess of things. Different checkpoints will affect this as well. I found that dpm anything, can provide the best results for trying to get the style right.
Art was trained on a class this time, class was design, where as previous versions were made without using a class token. Thought about trying stencil class instead but it isn't really close because there are multiple colors in the image, between 2-4 colors in dataset art.
Something to note that is useful for this model is if you are good at doing color separations for say, plating or vinyl application, the outputs from this model are near perfect for this use. Using something like Inkscape or Illustrator to extract the content for plating separations is great. I am referring to screen printing applications when talking about plating preparation.
Currently this model does what I want, to an extent, but it isn't great. I might revisit this one at a later time to help make it more versatile while maintaining style without nearly as much headache trying to control the style.
Different trigger: angst design
Best sampler scheduler: dpmpp_sde_gpu karras
Help push style by using negatives: photo, blurry, render
What's different?
Smaller dataset with flip augmentation
More responsive to prompts
Stiff in variety of outputs
Smaller file size for model
Trimmed down the dataset and cleaned up art via vector conversion. Edges are crisper, less colors, more separated. This version is kind of a side grade. It isn't exactly better than the previous or worse as a whole. Parts are better, like this version is more responsive, but it's also harder to push the style. The dataset has less variety so the model is a bit stiff. It can sometimes be impossible to get the style to push properly even with correct triggering in prompt. IMG2IMG is best for pushing the style on subjects that it just can't generate with prompting alone.
The sampler and scheduler can be really tricky with this model. With some scheduler sampler setups it will look like the model isn't doing anything at all or is just making really bad images, blurry mess of things. Different checkpoints will affect this as well. I found that dpm anything, can provide the best results for trying to get the style right.
Art was trained on a class this time, class was design, where as previous versions were made without using a class token. Thought about trying stencil class instead but it isn't really close because there are multiple colors in the image, between 2-4 colors in dataset art.
Something to note that is useful for this model is if you are good at doing color separations for say, plating or vinyl application, the outputs from this model are near perfect for this use. Using something like Inkscape or Illustrator to extract the content for plating separations is great. I am referring to screen printing applications when talking about plating preparation.
Currently this model does what I want, to an extent, but it isn't great. I might revisit this one at a later time to help make it more versatile while maintaining style without nearly as much headache trying to control the style.