You want more than a digital style - you want to feel brush strokes and see the built-up paint of an oil painting. You love physical objects and want your AI-generated art to feel like you're looking at a photograph of something analog, hanging on a wall somewhere.
This is the embedding for you. Download the 'classipeint.pt' file and trigger it in your prompt "art by classipeint" or "painted by classipeint" or simply "by classipeint"
It is reasonably flexible - I find I can prompt for fantasy elements, classic scenes, modern architecture ... it does sometimes take a little finessing but except for bad anatomy, I am using surprisingly few negative prompts.
Interested in generating your own embeddings? My Google doc walkthrough might help
You can rename the file and use that filename as the prompt. Just be sure your filename is unique and not something that may be an existing token that Stable Diffusion is trained on.
This embedding is a kind of companion to my first embedding, laxpeint, which provides a more digital-painting feel. Laxpeint is also more unruly and quirky (it often generates lots of baby heads? I don't know why) and this one suffers no such oddities that I've yet seen. They combine nicely together. Classipeint also does a good job combining with other embeddings.
Description
FAQ
Comments (17)
Lovely thanks! Not so many interested in painterly styles it seems but this exactly what I was looking for. Tried training using hypernetwork with marginal success cant seem get decent output with textural inversion at all. Any clues on settings for training would be much appreciated!
Glad you like it! Here's a Google doc walkthrough I made discussing how this embedding came together.
I'm really stumbling through a lot of this process myself, so don't consider me an expert by any means. But perhaps it's enough to get you on the right path? Let me know if you see some success and definitely point me to your embedding if you end up with one you like
Thanks so much! :)
This is definitely something I was looking for in a while, it's a shame it only works with SD 2.1, most models use previous versions like 1.5, any chance to get something done on an earlier version?
Much appreciated!
Sorry - I haven't shown much love to SD1.5 models since 2.1 came out (those extra pixels in the 768 model mean a lot!). But if I can carve out the time I'll see if I can train my dataset for it.
Please please make a version for SD1.5. Thank you so much
@barepixels I'm sorry, I just left SD1.5 behind. Especially with SDXL releasing within a month or so, the SD community should be 100% moving forward with that - it puts 1.5 and 2.1 to shame as a base model. Fine tunes of it will be absolutely amazing.
@eldritchadam honestly I can't wait to see more models begin to popup with SDXL, so far I don't seem to see a lot of 2.1 to begin with.
@eldritchadam Here is the problem, I really want to do a series of classical art nude painting, but 2.1 and SDXL censor stupid nipples. Perhaps you can share the 768 cropped tiles you used so I or I find someone to make for 1.5. I absolutely love you nude in the stream.
@barepixels sorry, I don't visit CIVITAI enough lately, didn't see your comment. Sure, I'll give you the images I trained on. Just a sec while I upload them somewhere ...
here you go - this Google drive folder has all the 768px crops I used and their captions. The Google doc above shares all the settings for the SD2.1 embedding. Hopefully it's pretty painless to adapt to a 1.5 training. Let me know if you see success with it!
@eldritchadam Thank you so much. BTW I recently installed "InvokeAi" for it's outpainting features. InvokeAi see and load your "laxpeintV2" BUT fail to recognize "laxpeint" and "classipeint" Anyway just reporting in. Thank you for the classipeint dataset
@barepixels ah, yeah - it's an oddity with Invoke. It's in there, but it shows up the the name I initially assigned it in training, hyphenated with the end training steps, so it looks like <adartaesthetic-198>
an unwieldy term to use, which is why I renamed it to classipeint. Didn't anticipate anyone would ever see my ugly 'adartaesthetic'
it's a mash-up of Adam D. Art Aesthetic. alas.
Great embedding!
Note that the embedding trigger for Invoke AI is <adartaesthetic-198>
Thanks! I only just installed invoke AI yesterday - it's fantastic!
I wasn't aware that the term I used in the TI training process embedded itself in the .pt file - but yeah, that was my training term before I renamed the file to upload. I thought all the SD GUIs simply called on the filename. There's probably a way to get the embedding itself to use 'classipeint' but I don't know how yet.
@eldritchadam strangely, invoke ai will use <(filename)> for the SD 1.5 embeddings, but for the 2.x embeddings, it doesnt seem to work, and the triggers inside only work (in my experience)
Can we get the same for SD 1.5, pleaaase? 🙏