DreamShaper XL Turbo in Fooocus with madebyollin/sdxl-vae-fp16-fix.
You can't set a default VAE to download, so I use this launch script (launch.sh
):
#!/usr/bin/env bash
set -e
export GRADIO_ANALYTICS_ENABLED=false
fooocus_dir="${BASH_SOURCE%/*}"
args='--always-high-vram --disable-preset-selection --disable-preset-download'
# honor preset
if [[ "$*" == *--preset* ]] ; then
args='--always-high-vram'
fi
# download fp16 VAE if flag is passed
if [[ "$*" == *--vae-in-fp16* ]] ; then
if [[ ! -f $fooocus_dir/models/vae/sdxl.vae.safetensors ]] ; then
echo "Downloading fp16 VAE..."
curl -fsSLo "${fooocus_dir}"/models/vae/sdxl.vae.safetensors https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/resolve/main/sdxl.vae.safetensors?download=true
fi
fi
# don't quote args so Bash word splits it
"${fooocus_dir}"/venv/bin/python launch.py $args "$@"
Guidance scale (cfg_scale
), sharpness, sampler, and scheduler are all overridden when using performance LoRAs (LCM, Lightning, Hyper). See async_worker.py
. Defaults are in config.py
.
This goes in presets/dreamshaper.json
:
{
"default_model": "DreamShaperXL_Turbo_v2_1.safetensors",
"default_vae": "sdxl.vae.safetensors",
"default_performance": "Hyper-SD",
"default_aspect_ratio": "1024*1024",
"default_image_number": 1,
"default_advanced_checkbox": true,
"default_styles": ["Fooocus V2"],
"default_loras": [
[true, "None", 1.0],
[true, "None", 1.0],
[true, "None", 1.0],
[true, "None", 1.0],
[true, "None", 1.0]
],
"checkpoint_downloads": {
"DreamShaperXL_Turbo_v2_1.safetensors": "https://huggingface.co/Lykon/dreamshaper-xl-v2-turbo/resolve/main/DreamShaperXL_Turbo_v2_1.safetensors"
}
}
Note that --all-in-fp16
applies to the U-Net and CLIP, not the VAE. You need to pass both if want everything in 16-bit. The flags are handled in model_management.py
:
chmod +x launch.sh
./launch.sh --all-in-fp16 --vae-in-fp16 --preset=dreamshaper