Optimum SDXL Usage

Here's a quick listing of things to tune for your setup:

Commandline arguments:

• Nvidia (12gb+) --xformers
• Nvidia (8gb) --medvram-sdxl --xformers
• Nvidia (4gb) --lowvram --xformers
• AMD (4gb) --lowvram --opt-sub-quad-attention + TAESD in settings

Both rocm and directml will generate at least 1024x1024 pictures at fp16. However, at full precision, the model fails to load (into 4gb). If your card needs --no-half, try enabling --upcast-sampling instead.

System:

• (Windows) Downgrade Nvidia drivers to 531 or lower. New drivers cause extreme slowdowns on Windows when generating large images towards your card's maximum vram.
  This important issue is discussed here and in #11063.
  Symptoms:

  • You see Shared GPU memory usage filling up in Task Manager
  • Your generations that usually take 1-2 min, take 7+ min
  • low vram cards are generating very slowly

• Add a pagefile to prevent failure loading weights due to low RAM.

• (Linux) install tcmalloc, greatly reducing RAM usage: sudo apt install --no-install-recommends google-perftools (#10117).

• Use an SSD for faster load time, especially if a pagefile is required.

Model weights:

• Use sdxl-vae-fp16-fix, a VAE that will not need to run in fp32, for increased speed and less VRAM usage.
• Use TAESD; a VAE that uses drastically less vram at the cost of some quality.