Better default scheduler and number of inference steps

[Necr0x0Der]

Some experiments should be conducted to optimize the number of inference steps used by default together with the default scheduler.

[noskill]

I would just take settings from showcase images from dreemshaper or "I Can't Believe It's Not Photography" on civitai.

DPM++ 2M Karras is often used on civitai, diffusers implement it as sample_dpmpp_2m but it doesn't work with custom pipeline like long prompt weighting. I suggest we use DPMSolverMultistepScheduler since it gives most similiar results to sample_dpmpp_2m. "ICBINP" showcase images use on average 25 steps and dreamshaper 28. Almost all images have clip_skip=2

[noskill]

For SDXL they often use 50-100 steps, but there are good images with 20-25 steps

[Necr0x0Der]

I'd prefer to perform some tests. Also, some examples of different control net pipelines explicitly specify other schedulers. I wonder if there is a reason behind this.

[noskill]

@Necr0x0Der what would be score function in such tests? In civit.ai we have likes and downloads which is good proxy for image quality.

[noskill]

we can adapt callbacks to turn off guidance at first few steps of diffusion huggingface/diffusers#7761