This is an embedding to provide a loose digital painting style
I highly recommend the new V2 embedding - it is much more flexible, with far fewer odd artifacts. V1 had a heavy leaning toward medieval classical themes and artifacts that required extensive positive and negative prompts to steer properly. V2 (generally) just gives you what you ask for.
One quirk, and I love it so much, is for some reason when I prompt for a skydiver, I routinely get a skydiver on a horse. No horses were included in the training images ... so how that snuck in is a mystery
Interested in generating your own embeddings? My Google doc walkthrough might help
-- -- --
After testing prompts with this embedding a few more days, I've found a fascinating tweak on its style can be to include prompt terms like "anime" and "cel-shading" to flatten some of the rendering down. It won't actually get you a truly anime look. It does get you a more graphic, punchy illustrative feel, but still with really rich coloring. Throwing other aesthetic terms and artist names at this embedding can move it in interesting directions.
Description
added some new images to train anew with a better understanding of the potential impact of good captions and initialization text - much more flexible and easy to use embedding than V1
FAQ
Comments (6)
Nice work!
Do you think you could bring this to SD 2.1? There's a bad hands dataset, and a good hands dataset.
https://civitai.com/models/4586/bad-hands-training-data
I'm afraid I don't follow your question. The embedding was trained on and SD 2.1 and will only work on SD2.0 or SD2.1 ... how would I further bring it to SD 2.1?
Also, I don't know what you would want me to do with this embedding in relation to someone's dataset?
No, that's not what I meant. Sorry, I can see how I wasn't clear before.
What I mean to say is that I like this, and your other work, and I saw your document going through the steps to create TI embeddings. And I wondered if you might have a go at bringing the Negative/Positive Hands embeddings to SD 2.1. For some reason the A1111 WebUI gives me black images when I run it on the machine I typically dedicate to this kind of work, but I might try it on my main desktop, since it seems it will be fine for creating embeddings on a smaller GPU.
@hughph I understand you now! Yeah, embeddings can only be used with the model they're trained on so if nobody's made a good negative hands embedding for SD2.1, seems like it needs to be made! I'll see if I can get my head around it. I haven't really put much thought into that aspect of embeddings, just putting most of my thoughts toward art styles, but I might be able to do something there.
v2 gives too many random dwarf faces, why?
honestly, it's just been too long since I've used this embedding - or SD2.1 at all for that matter. I can't provide any insight here. Sorry,