Private commission on fiverr
Get early access to my upcoming NSFW Lora on my Patreon.
early lora share on discord for patreon.
I officially host and maintain all my models on Tensor.art
I was hoping you could support my work by joining any one of them and get early access to all my upcoming loras and other perks, such as fan requests and a Discord role.
Join my Discord Server
check the images for prompts
use lora at 0.7-1
Adetailer for faces
Img2img upscale
4x-ultra sharp
comment you idea or request
Description
FAQ
Comments (10)
Hey. At this point we might be making a few too many models of the same style 😅
Now we have 6 models where the most downloaded is pretty mediocre. Ours are close to indistinguishable.
I was wondering, do you expect any concrete improvements with a LoCon? I trained just that and it can't exactly compensate for the large variance in ratatatat74s art.
i will test my lycoris model soon. i think I got pretty close to the original art style. lykons model is really old and it is one of the very early Lora models in Civitai so it might look low quality. i already made one lycoris with the same dataset. now i need to test it. im hopeful lycoris will give a better style
More is better. This style seems really hard to replicate in AI so it's cool to see the variation and progression among different people.
@oceanic if only most of them weren't completely under- or overbaked. I regularly find models that barely have any resemblance to the style they are supposed to replicate. Although it isn't exactly difficult to get close.
For ratatatat74 getting all the nuances would probably involve a huge dataset with proper tagging to differentiate between the manga/pinup/colorized images. I'm not sure it is worth the effort considering that it would drown in the "junk".
I also think it is really interesting to see the progression but in some sense seniority prevails - the model that can occassionally get it right and has the most downloads will always be anyones first choice.
@StickyRicky Well, I highly disagree but to each their own
so what is you suggestion? @StickyRicky
i normally dont search for models as if i need anything i would make it myself . and i think do try to maintain a good quality when i upload any lora. and i dont think it would be nice to stop others from trying. you might be upset( im not sure) as the highest downloaded model for this style is not representing the style properly but and you are saying mine and yours looks very close but you need to understand that lykons model is the oldest and he is also the top creator and early uploader so his model will get more download thats natural as when he uploaded that model me or you were not there to upload anything our own.
i dont make lora for downloads i make what i like doesn't matter how popular that concept is. like my 'free style' lora is not getting much downloads compared to this model but still i like the free style more more and i will use that model myself more,
@Goofy_Ai I really came on a bit too agressive, didn't I, venting my frustration.
I certainly don't have an issue with your model quality. I have just spent so much time on sorting through different - barely working - models before I started making them myself.
I wish I had a viable suggestion. In theory you would need (many) reviews that do not start at 4/5 as these should be a much better metric than downloads. Currently almost all reviews are 5* which isn't helpful to the uploader or other users.
While anyone should be able to train, test and upload models, you only make improvements through critically judging them. Give me twice the models for each style or character and proper reviews and I'd be happy. This would bring the best techniques to the forefront which would be the best for everyone.
(Yep, that's an utopic notion)
On the subject of Lykon, I understand that the model is pretty old and have no beef with them. This was just an example of an old model - which tends to mean that they are worse than what we do now - being the first thing people download. Speaking out of experience because I was one of them and really struggled getting good results.
@Goofy_Ai Anyways the LoCon turned out great ^^. Could you share your learning rate/image*repeats per gen that you use? Did you manually tag the images or use automatic tagging and added additional prompts?
@StickyRicky for styles i normally auto tag them and . while collecting data i try to limit watermark as much as i can. for cutesexyrobutts v2 i manually removed 400+ watermark. and for setting check the metadata as i dont have any fixed config. i change my settings depending on the concept and image amount.