If you like my contributions to the community, please consider Buying Me A Coffee or check out my other sites on AllMyLinks
TRAINING INFO
Trained on 47 images
Trained at 1024 resolution
RECOMMENDED SETTINGS
TXT2IMG - A weight of 1 works just fine and will yield some interesting results, but I recommend a weight of 0.4-0.7 for more flexibility while still keeping Bekinski's style.
PROMPTS - I highly recommend using prompts that already describe the style of Bekinski's work, including but not limited to: fog, structure, ominous, brown, sky, tombstones, tissue, decaying, backlit, bright, cracked, mysterious figure, crucifix, towering, silhouette, cliffs, orb, pillars, glowing, floating, massive, pathway, walls, gateway, portal, fossils, cloak, hooded, monolith, moon. Including prompts like these will give the best results. For a full list of the prompts/captions used for training, check the Lora file's metadata in your webgui.
Description
Same dataset but refined the captions so the prompts should work better and be more precise
FAQ
Comments (17)
Very nice, I wanted to do it but was too lazy. Did you notice there is a website that has 600+ works of his? If you're interested maybe we can collaborate on it and train something on that big dataset
I didn’t know about that website. What’s the link?
Zdzislaw Beksinski - 708 artworks - painting (wikiart.org)
708 artworks according to the website, though I suspect there might be some duplicates
@Atomease Whoa 😳 that’s awesome thanks!
at the moment I'm trying to train LoRA from 535 pictures. but with a size of 512x512. I started at first at a size of 1024x1024, but the estimated time was almost 70 hours))) it's too long)
@garand343635 😆 what’s your gpu? I’ve got a 4090. If there’s a way to share the credit that you’re happy with, I’d be glad to train it with your dataset. I’m not too familiar with how credits for collaborations work on here or if there’s a way to do that
@rjox I have a regular 3060. and there are a couple of hours left until the end of the training) 4 hours have already passed)
@garand343635 ah cool. looking forward to seeing it!
@rjox let's see what happens there)) and then maybe there will be some garbage)
turned off the electricity in the area... all the data flew away... will have to start over...
@garand343635 NOOOOOO. let me know if you want me to train it
@garand343635 use google collab , i have a link if you like
@GG_System it does not suit me.
There was some sort of power failure. was not 3 hours. but as it appeared, I set the training for the night. is already ending.
the training is over. I'm trying to do something
I have already published my model. but I will also use this one, because the results are completely different. on yours, I can just slightly stylize some person, for example, and he is well recognizable. and in my model, even if you give a small weight, the influence of the author is still very strong))) because there were a lot of his works, and the picture in any case becomes more abstract and surreal
ahh come on guys now I have 2 more Beksinski to review as soon as the day after ? 😂