CLiP Skip
[!WARNING] Using CLiP skip values without understanding their implications can lead to unexpected and most likely corrupt image outputs.
What is a CLiP skip
The CLiP text encoder model consists of multiple layers, each providing a different level of specificity
Lower Clip Skip values mean that more layers are used, resulting in images that closely match the prompts
Higher values skip some of these layers, leading to potentially more creative outputs but with less adherence to the original prompt
Why did it became popular? Because original NovelAI SD 1.5 model was trained with CLiP skip value of 2
And that model was used as basis for many other anime-style models, leading to a common practice of using CLiP skip value of 2 for anime content
However, anything newer than that, like pretty much ANY SDXL or FLUX or other recent model,
is trained with CLiP skip value of 1, and using CLiP skip value of 2 on these models will lead to bad results
However, there are still many recommendations on the internet to use CLiP skip value of 2 - and most of them are outdated and wrong (unless we're talking about NAI SD 1.5 derivatives) The problem arises from the fact that many SD applications such as original Automatic1111 do not even implement CLiP skip for SDXL and other newer models, so setting it to 2 does not change anything and does not lead to any issues
SD.Next, on the other hand, does implement CLiP skip for all models, so setting it to 2 will lead to bad results with SDXL and other newer models
If you want to experiment with CLiP skip values, you can do so, but be aware of the implications
To change the CLiP skip value, you need to unlock it in settings and then set it to the desired value
Note
SD.Next also supports non-integer CLiP skip values, such as 1.2, 1.5, etc.