How does the "Clip Skip 1" value works?

[hansolocambo]

I'm used to train Clip Skip 1 for SDXL and Clip Sklip 2 for PonyXL. But a tutorial (latest ESCourses on Youtube about OneTrainer) showed me a "Clip Skip 1" of 0 to train ... what I would have thought should have been 1 ? This line in the UI being called "Clip Skip 1" makes it quiet confusing (to most average users I guess).

Which "Clip Skip 1" value should I use ?

• For SDXL I want Clip Skip 1, so should I use a value of 0 like in the tutorial I saw ? Or 1 ?
• For a Pony XL training (I want Clip Skip 2), should I use a value of 1 ? .... or 2 ?

I could really use some help on this one.
Thanks.

[Davikar]

I think Clip Skip 0 is the same as 1 (0 is off, which means 1) and Clip Skip 2 would be 2.
At least that's what I've been assuming...

[Nerogar]

Clip Skip = 0 means that 0 layers are skipped. The output of the last layer is used as input into the UNet
Clip Skip = 1 means that 1 layer is skipped. The output of the second to last layer is used as input into the UNet

I don't know what exact naming some of the other tools are using, but using Clip Skip 2 as the name for the second to last layer seems very confusing to me. That's why I didn't use that naming.

[hansolocambo]

Well thanks for the answer but it doesn't really. I'm not asking for coders explanations, but for dumb artist explanations. Please try to consider that average users don't have your brains.

"using Clip Skip 2 as the name for the second to last layer seems very confusing to me"
What seems confusing to you is clear to everybody else using CivitAI for example... You chose another way of writing all that, well fine. And your way is better than everybody, fine too. But why not make a UI that people understand because it's compatible to the terminology we see everywhere else. Then in the background your app does what has to be done. But we don't have to see it or to know what it means. When people sculpt with their stylus in ZBrush they don't care about the technical things happening in the background. Simple, intuitive, that's what end users need sadly. Nothing more ;)

At the end, thanks for you answer BUT : "means that 0 layers are skipped. The output of the last layer is used as input into the UNet": this doesn't mean shit to my brains, and I'd guess it's the same for 99% of SD users.

So please, instead of words most users cannot understand. Just try ;) to make it simple.

[O-J1]

You are wrong. Stop coping. Bad standards should not be used for conviences sake. That thinking is why we have QWERTY a keyboard layout that not only intentionally slows you down but also introduces additional finger stresses leading to potential RSI.

[hansolocambo]

"you are wrong".
Wow that's super helpful. Thanks for that. Now I really know what I should do...

[mx]

If you're coming from Automatic1111: take the value you'd use for clip skip there, and subtract one from it, and that's what you'd use in OneTrainer.