sd-tools

pip install sd-tools

Basics

sdxl 'locomotive comming' --size 512

If the model can't be downloaded, try using a huggingface mirror:

export HF_MIRROR=https://hf-mirror.com

To get better result, use more steps and higher resolution:

sdxl 'locomotive comming' --size 720 --steps 4

Generate more with different resolution:

sdxl 'locomotive comming' --size 1024x576 --steps 4 --count 4

Interactive Mode

The loading of the model can task some time, use -i to enter interactive mode, keeping the model loaded:

sdxl 'locomotive comming' --size 1024x576 --steps 4 -i

Edited prompt and press enter:

> locomotive comming

size, count and cfg can also be set on the fly:

> locomotive comming size=1024

Custom Models

To use a specific model:

sdxl --model Lykon/dreamshaper-xl-lightning \
 --steps 6 \
 --size 1024x576 \
 --scheduler 'DPM++ SDE Karras' \
 --cfg 2 \
 'locomotive comming'

More models can be found on huggingface and civitai, models with Lightning or Turbo can generate image in less than 8 steps.

For models without fp16 variant and safetensor format, add --no-fp16 and --no-safetensor

sdxl --model RunDiffusion/Juggernaut-XL-Lightning \
 --cfg 1.5 \
 --steps 6 \
 --size 832x1216 \
 --scheduler 'DPM++ SDE' \
 --no-fp16 --no-safetensor \
 'locomotive comming'

To use models downloaded from civitai:

sdxl --model ./model.safetensors 'locomotive comming'

Loras

Loras are like plugins for the base model, multiple loras can be used:

sdxl \
 --model SG161222/RealVisXL_V4.0_Lightning \
 --steps 8 \
 --size 1024x576 \
 --loras ./lora1.safetensors ./lora2.safetensors:0.8 \
 'locomotive comming'

Speed Up Generation

Using --lcm, --tcd or --lightning to speed up generation, not necessary for turbo/lightning models:

sdxl \
 --model SG161222/RealVisXL_V4.0 \
 --steps 4 \
 --size 1024x576 \
 --tcd 1 \
 'locomotive comming'

More Controlling

Canny

sdxl 'protrait of a man' --canny photo.jpg --canny-low 100 --canny-high 200

Depth

sdxl 'protrait of a man' --depth photo.jpg

Pose

sdxl 'protrait of a man' --pose photo.jpg

Style and Face

Photo Maker

pip install sd-tools[photomaker]

use img to indicate the reference target

sdxl \
 --model Lykon/dreamshaper-xl-lightning \
 --steps 8 \
 --size 1024 \
 --scheduler 'DPM++ SDE Karras' \
 --photo-maker ./photos \
 'portrait of a man img'

reference images can also be specified by filenames

--photo-maker 1.png 2.png

IP-Adapter Plus

style reference

sdxl 'Mario' --ipa-plus ./styles/

face reference

sdxl 'Mario' --ipa-plus-face face1.png face2.png

combined

sdxl 'Mario' --ipa-plus ./styles --ipa-plus-face ./faces

IP-Adapter FaceID Plus

pip install sd-tools[faceid]

sdxl 'Mario' --ipa-faceid face.png

InstantID

pip install sd-tools[faceid]

sdxl 'Mario' --instanceid face.png

HTTP Server

sdxl 'Mario' --listen '127.0.0.1:8800'

invoke using curl:

curl '127.0.0.1:8800?prompt=shell&seed=1' > image.webp

More Examples

sdxl --model RunDiffusion/Juggernaut-XL-Lightning \
 --prompt 'portrait of a woman' \
 --cfg 1.5 \
 --steps 5 \
 --size 832x1216 \
 --scheduler 'DPM++ SDE' \
 --no-fp16 --no-safetensor

SD 1.5

TBD