Type | |
Stats | 420 621 |
Reviews | (73) |
Published | Apr 13, 2024 |
Base Model | |
Training | Steps: 1,000 |
Trigger Words | stardeaf_darkcore |
Hash | AutoV2 E0B2F19F97 |
This is art style embedding. It creates comic-book-like inked illustrations, dark, gloomy and simplistic with lots of sharp angles. You can guess the inspiration here.
Notes and suggestions:
It works well with some checkpoints and not so well with other. Experiment. Realistic models seem to work best. Or look at suggested resources below, these are ones I was using for sample images.
Don't use any quality tokens, "best quality", "masterpiece" and such. Also don't use negative prompt. Likely it will kill the style.
Weight the embedding against the rest of the prompt. Like, give it 0.7 and the rest of the prompt 0.7 too and see the result and then try to increase/decrease the weight of style token.
Sample images were created in InvokeAI, it has slightly different grammar than A1111. E.g. (crackled eggs)0.5 - everything in brackets gets 0.5 weight.
The embedding has 4 vectors. Training took about 1000 steps. I did it in passes, 600 first then checking results, adjusting strength and adding several 100 step passes.
Have fun.