LoRA weight 0.6 - 0.8 are better.
v2 improvement, reduced multiple tails issue.
v2.1 Experimental "no tail" edition. The "no tail" effect is uncertain, about 60% will have a tail added.
V2.2 For Model & VAE compatibility test version. (Maybe delete it later?)
I made it without a trigger. but prompting for anthro, anthromorph, shark, shark_girl, will good results.
I use an online service to generate these images, so the metadata seems to be dropped sometimes. sorry for any inconvenience.
Description
LoRA weight 0.6 - 0.8 are better.
v2 improvement, reduced multiple tails issue reduce.
FAQ
Comments (16)
What checkpoints do you advise using this with?
I mainly use this. Try it.
(furry & kemono mix) https://civitai.com/models/90762
@tommyderhund Appreciate it :D
What is the trigger word?
I made it without a trigger. but prompting for anthro, anthromorph, shark, shark_girl, will good results.
@tommyderhund Thank you!
Cool lora and pictures, but I can't even repeat them roughly, and the metadata is partially broken. Please share an image with full metadata, if possible.
Thank you fp16_guy. I respect your work. and sorry. I'm using a small poor laptop and online service to generate LoRA and images. so I can't seem to include full metadata in the image file.
@tommyderhund, oh, I get it now, thanks anyway! And I periodically use Colab, because I also have my own small poor laptop. If you ever decide to try it too, be sure to update your previews, the world (and I) need girls with sharp teeth and a gorgeous figure!
As for me, I do a lot of photoshopping and tweaking and feeding back into img2img, which is why my stuff frequently doesn't have metadata, because the initial txt2img is often fairly far removed from the final product, which is itself often a photoshopped together amalgamation of like 5 different img2img's.
@neonmodder Agree. "photoshop + img2img" process is helpful sometimes.
In your examples, I do not see where the Lora is used (e.g. <lora:anthro_shark_girl_v2:1>).
Can you suggest why?
I use an online service (eg, PixiAI, SeaArt) to generate the images, so the metadata seems to be dropped. LoRA generation also uses https://civitai.com/models/22530/guide-make-your-own-loras-easy-and-free. For more information read the discussion with fp16_gui on this page. A strength of 0.6 to 0.8 is recommended.
@LindezaBlue use and see.
@LindezaBlue I can tell you that it does better than without Lora
@Sylvanal In early days, I read metadata contained in png file as a txt file and copied and pasted it on the upload page, but it's not fun and I don't want to do it anymore. If you have any questions, I will answer them as much as possible. Please forgive me for any inconvenience. (Even my hard-to-understand machine-translated English)