A fairly standard cute anime style, but with exceptionally sharp image quality and unusually good hands. This might serve as a good basis for other Lyco mixes.
Note that this model is unusually large for a lycoris, but I typically run a resizing algorithm on it that attempts to preserve as much good information as possible, and that didn't shrink it very much, which indicates to me that the higher quality is tied to the size of the file. I'm not going to be putting out a smaller version of this one (although if someone else does, I'll link to it from here).
Description
FAQ
Comments (15)
This one has a pretty neat soft-color style.
It's cute indeed..
I challenge you to go even cuter! Bigger eyes, borderline chibi look! I'm working on something like that myself, but needs a lot more work.
I'd like to know how you "compress" the LyCORIS like this (and how you got a file so big in the first place - much higher than 128/128 dim and alpha? Or was this an extraction from a checkpoint?)
I use kohya_ss to compress it. I set the rank to 128, the dynamic method to sv_fro, the dynamic parameter to 0.98, and the save precision to fp16. I don't notice much in the way of quality loss, and often times it will reduce a lora from hundreds of megabytes to tens.
Anyway, in my experience, a float precision locon with 128 dimensions starts of at about 600 megabytes. Loras are a lot smaller, I believe.
@_Envy_ Thanks for your reply.
@datSato Also... challenge accepted.
What "resizing algorithm" are you using? Maybe it can be modified
sv_fro with a dynamic parameter of 0.98.
@_Envy_ I'll take a look, I'm actually working on some other tests with the LoRAs code, I'll see if I find anything useful
@_Envy_ from the few tests so far, dynamic_method="sv_fro" and dynamic_param=0.98 doesn't reduce the size too much, since from what I understand of the functions, it's keeping 98% of the values of the S diagonal. It also doesn't change the results much, but in that case it might not be worth doing the resize. I'm now doing some tests with 0.5 and will test the other options to see if I find anything else useful.
Some tests here: https://civitai.com/posts/289182 . Still testing more ideas to see if it's possible to get closer to the original LoRA results with smaller LoRAs
More tests here: https://civitai.com/posts/289531 learned a lot already, will continue testing more and I made 3 new functions to test besides sv_fro and the other 3 that are available by default. There are interesting opportunities to reduce the LoRA sizes :D
@victorc25744 Those are good, but the key thing I want to preserve is hand quality. For whatever reason, this lora makes really good hands.
@_Envy_ I'll try to find out what exactly part of the LoRA does the hands, maybe I can get a small LoRA just for that and it can be reused elsewhere. Out of curiosity, was it merged or trained with the same or part of the dataset of your hands LoRA?
@victorc25744 No, the hands lora isn't part of it.
@_Envy_ Nice. I don't know if it will be interesting enough, but I managed to make an 8,7MB LoRA out of this LoRA, by keeping only the layers that seemed to help with the hands. There are small changes in other parts of the image, but they are often very small. I wanted to try making something that could be complimentary to the GoodHands LoRA, here is a test with it: https://civitai.com/posts/292569
Next weekend I'll try doing some more tests with one of your LoRA series to see what else I can learn, this was very useful for me :D
only thing annoying from these is having to manually updating the preview image
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.