/h/anime
TouchFluffyTails Them/They https://x.com/AyakonArts :marseydab2:  8mo ago (image post) 431 thread views #255659
Reported by:

Rabbit hole miku lora

Unable to load image

https://civitai.com/models/363402/rabbit-hole-hatsune-mikupure-pure

idk where to put it but here you go if you want to make images of the bunny

https://i.rdrama.net/images/17111524182631443.webp https://i.rdrama.net/images/1711152419207502.webp https://i.rdrama.net/images/17111524202307024.webp https://i.rdrama.net/images/17111524210696306.webp https://i.rdrama.net/images/171115242191246.webp https://i.rdrama.net/images/17111524228952348.webp https://i.rdrama.net/images/17111524237968242.webp

24
New Old Top Bottom Controversial Random
Jump in the discussion.

No email address required.

Really wish we could speed up this timeline and get to the part where you get thrown in prison for child porn.

Snapshots:

https://civitai.com/models/363402/rabbit-hole-hatsune-mikupure-pure:

Jump in the discussion.

No email address required.

jesus christ snappy

Jump in the discussion.

No email address required.

holy snappilent

Jump in the discussion.

No email address required.

:#marseyflushzoom:

Jump in the discussion.

No email address required.

good, finally something not early acc-

>270mb char lora

:#marseyreluctant:

why would you make it larger than the dataset it's representing (especially in latent space, ain't no way you have 270mb of latents)

the only thing that does is make it less compatible with other loras that may also have their problems

https://i.rdrama.net/images/17111568166560113.webp

Jump in the discussion.

No email address required.

Because memory is cheap and there is no reason to bother making it smaller - I train at 128 dim

Sdxl Loras get big fast for quality

Jump in the discussion.

No email address required.

no reason to bother making it smaller

Preventing overfitting which makes it behave like crap on other models or with other loras

This is a relatively simple char lora, not one of those loras with hundred or thousands of styles inside

Jump in the discussion.

No email address required.

On sdxl 128 dim I have found to be a sweet spot, trying to do 256 or higher is where issues like that start (1.5 worked best on 256)

Jump in the discussion.

No email address required.

I'm always impressed with your AI posts.

Jump in the discussion.

No email address required.

:marseyhearts:

Jump in the discussion.

No email address required.

@ChristMAsLLOWEENewYear

Jump in the discussion.

No email address required.

:#marseycontemplate:

Jump in the discussion.

No email address required.

Getting back into running SD at home, could I ask why I keep on getting this error message?

A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.
Jump in the discussion.

No email address required.

Nevermind! Got it working!

Jump in the discussion.

No email address required.

Link copied to clipboard
Action successful!
Error, please refresh the page and try again.