Better AI Hands. How to Install Smea Dyn for Stable Diffusion.

preview_player
Показать описание


Рекомендации по теме
Комментарии
Автор

Since the name Euler comes from the mathematician Leonhard Euler, who was born in Switzerland in the 18th century, I can confirm as a native Swiss that it is pronounced like "Oiler." This is a well-known family name here. That should end the discussion once and for all, at least for those who read this. ;-)

subsudo
Автор

Creating a sampler for dealing with the hand problem is just a mere band aid. The issue is simply that hands are a complex concept and have a much higher variety relative to the rest of the body. The same goes for faces in images where the face is not the focus. But hands are way more difficult. Image augmentation during training/fine tuning is a better approach on the long run. But its difficult. The amount of samples and the right balance means a shxxx ton of work.
Another issue are relative low res samples. For a model trained with samples below 1mp it is very difficult to pick up the details of hands. And lets not get started about the "right" tagging.

madrooky
Автор

Euler was a Swiss mathematician born in Basel - the German-speaking part of Switzerland.
In German, the name is definitely pronounced like 'oiler' - same as Europa is pronounced 'oirohpa'.

Steamrick
Автор

It is "Oiler", you're correct. I've been dealing with Euler rotations for over 20 years.

nebuchadnezzar
Автор

Dead jokes can also be stored with your mummy.

andyrevo
Автор

It might be good for inpainting just for the bad limbs. Thanks for the video!

skycladsquirrel
Автор

Great vid as always! Your examples focused on Anime, any tests against photoreal and it's impact on both SD15 and SDXL?

GrocksterRox
Автор

Would be interesting to try things like "hand shaking" or "holding hands" or "hugging". It seems like those images always confuse the models.

KDawg
Автор

It helps me, you’re great, I will try it, thanks man

popular
Автор

same sampler is in Comfyui - install it via manager as node

NotThatOlivia
Автор

This is awesome! available for comfuyi as well?

blademarketing
Автор

Is there a scenario for Heun samplers? I always get very subpar results with Heun when testing pretty much anything. Maybe it is better for cartoons?

sandnerdaniel
Автор

I got a runtime error: the size of tensor a (120) must match the size of tensor b (96) at non-singleton dimension3. when trying to use Euler SMEA Dy at the sampling method. any clue how to fix it this problem?

drericclinic
Автор

I just tried them all with a selection of models (1.5 and XL) using the prompt "normal women looking at their normal hands". The results were hilariously terrible.

frankhall
Автор

No issues about how you say "Eular", but "centaur" on the other hand... ;). (in english it's sen-tore rather than ken-tore)

SyntheticSurrealism
Автор

Getting an error when using it :( - The size of tensor a (80) must match the size of tensor b (64) at non-singleton dimension 3

danieldimitrovdrums
Автор

the reason hands are so difficult for AI is proof the technology was created by aliens

CoconutPete
Автор

didn't control net already solve the hand issue ? or is this method more data efficient or have other benefit above controlnet

keenheat
Автор

What if you run another pass after dpm karras using this new model?

Brandon-xvk
Автор

how to install Sampler: DPM++ 2M Karrass .... i have AMD graphic card, i want to generate image to image.

createfun