A rather dark-psy inspired model (over)trained on 200+ AI Generated training images and a custom set of AI Generated regularization images, for a total of 23K steps.
The model was trained on images generated with SD 1.5 using the following artists and artstyles as sole prompts.
Pablo Picasso, Alex Grey, Zdzisław Beksiński, Katsuhiro Otomo, H.P. Lovecraft, Clive Barker, M. C. Escher, H.R. Giger, Yves Tanguy, Andy Warhol, Jackson Pollock, Frida Kahlo, Marcel Duchamp, Salvador Dalí, Georgia O'Keeffe, Yayoi Kusama, Jean-Michel Basquiat, Ai Weiwei, Dadaism, Fluxus, Surrealism, Abstract Expressionism, Pop Art, Outsider Art, Neo-Dada, Street Art/Graffiti, Installation Art, Performance Art, Biomechanical, Soviet Propaganda
Even the regularization images are AI generated interpretations of those artists/art styles!
Use this model if you want to get surprises. Even just prompting for the class and keywords (`artstyle bse`) will get you quite some results.
It cannot do faces, hands very well. You can get lucky (see example shots), but it is not the purpose of the model anyway.
I strongly suggest using this model when you want to get very artsy, psy, otherworldly.
I did a lot of experiments with this model and all its intermediaries (I saved each 5k steps during the dreambooth training) and found that it works best with the following settings:
Positive prompt:
artstyle bse ((YOUR TOPIC GOES HERE)), best quality, masterpiece, ultra high res, (photorealistic:1.4), raw photo, sharp focus, HDR, detailed skin
Negative Prompt
Anime, cartoon, ng_deepnegative_v1_75t, (badhandv4:1.2), (worst quality:2), (low quality:2), (normal quality:2), lowres, bad anatomy, bad hands, normal quality, (monochrome), (grayscale), watermark
Resolution:
512^2
Sampling method
DDIM
You can also get interesting results with larger resolutions like 1200x900
, by increasing the steps massively, although not useful for subjects, it can be interesting for "backgrounds"
Resources:
I used both badhandv4
(https://civitai.com/models/16993/badhandv4-animeillustdiffusion) and ng_deepnegative_v1_75
(https://civitai.com/models/4629/deep-negative-v1x) in almost all prompts.
PS:
you will notice that the prompts in all pictures say the model used is baked
. That is because I dumbass renamed the model after I downloaded it from my SD online instance and after I generated all tons of images with it. As I had named it baked
in the SD instance, that is what the images meta data tell too. However, the model used to generate those image is the current uploaded TGL Psy
model, you can verify this by simply recreating the images.