Skip to content

Optimum SDXL Usage

ClashSAN edited this page Dec 15, 2023 · 6 revisions

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. If your AMD card needs --no-half, try enabling --upcast-sampling instead, as full precision sdxl is too large to fit on 4gb.

System:

  • (Linux) install tcmalloc, greatly reducing RAM usage: sudo apt install --no-install-recommends google-perftools (#10117).
  • Add a pagefile/swapfile to prevent failure loading weights due to low RAM.
  • Use an SSD for faster load time, especially if a pagefile is required.
  • Have at least 24gb ram on Windows 11, and at least 16gb on Windows 10

Model weights:

  • Use sdxl-vae-fp16-fix; a VAE that will not need to run in fp32. This will increase speed and lessen VRAM usage at almost no quality loss.
  • Use TAESD; a VAE that uses drastically less vram at the cost of some quality.
Clone this wiki locally