Doesn't make images realistic, but enhances the realism of images - it works well with other lora that I've tried so hopefully it's helpful. All examples are raw, unedited outputs.
V2.0:
Revised dataset - more NSFW - trained much longer. This tends to have a more noticeable impact on color temperature (warmer) and composition compared with other versions, but I also think it's more realistic
V1.0:
Original realism enhancer that I made that seemed to work reasonably well and maintained consistency with loras.
I've continued to experiment and will release some updates - unless there is a meaningfully better model, I'll keep iterating under the V1.0 enumeration.
ALT 1.0:
Alternative dataset from V1.0 - not better or worse, but it seemed usable so I figured I'd share it - I think it's close to 2.5D than V1.0
ALT 1.1:
Alternative training optimizer with the same dataset as ALT 1.0 - no better or worse
Description
Adjusted the dataset - not necessarily better or worse - but usable and figured I'd share it
FAQ
Comments (9)
This is amazing!
Is it working well for you? I realize that I was too busy testing it with comparison prompts - not removing the style or artist tags or anything and only applying my lora - I never actually tried just making a realistic image with it...
@CallMeMaybe its working great. im using another lora together with this and getting some nice realistic images. I was testing lots of different lora cominations and so far your lora gives best results.
ALT 1.1. does a great job of maintaining the artist's style but still apply a realistic filter to it! Only thing that stands out a lot to me is the hair not being that realistic? Like, it works if you don't tag a specific anime character, then it looks more like natural hair, but if you do it looks more like a styled wig, y'know? Either way, nice models!
Thanks a lot! Hair has been a big struggle...the way I'm making the dataset is that I'm taking original anima images - keeping the prompt the same - and passing it through my personal 'anime 2 real' workflow, then training on those images. I'm trying various combinations of lora and models for making the dataset images more 'real' but hair is really really difficult to get to look realistic while maintaining the core anime style/details
the ALT dataset I tried making it less 'real' to preserve more of the anime details, but that means the hair is even less real.
I'd love if you could post some of your generations with my lora!
@CallMeMaybe I have been looking for a long time for a replacement for what was once the 1.5 SD hairdetailer. Most solutions for Illust have been struggling with these same issues. I'll keep trying
If you want photorealistic images, meaning photography, you dont use the "realism" or "photorealistic" tags. Those just make the image more 3d.
To get photo, use "real life photography" ideally with high weights like (real life photography:3)
Thanks! Yes the idea with this wasn't to try to make 'photorealistic' but to just enhance realism from the 2d aesthetic, resulting in a 2.5D or 3D kind of style.
@CallMeMaybe gotcha