Stable Diffusion

GPU Capable

Easy Diffusion

By far the easiest way to install Stable Diffusion is via this amazing GitHub repo at https://github.com/cmdr2/stable-diffusion-ui which has implemented a "One-Click" solution where you download some files, extract and run. There is another link to the same resources at https://stable-diffusion-ui.github.io/, now known as Easy Diffusion

This solution provides a very capable Browser based UI called Easy Diffusion, which can also be configured to be accessible from your local network.

It can also be configured to run on CPU only if your GPU is not compatible, but obviously the rendering can then take minutes instead of seconds.

The above allows significant configuration including different models, LoRA tuning etc, try it out.

Add On's and Features

Checkpoint (pretrained model with set of knowledge) -- provides knowledge of what things look like in general e.g. Epic Realism
LoRA (Low-Rank Adaptation) -- provides an easy and low resource way to tune a model to be able to do some very specific output such as a character, uniform or equipment e.g. Star Trek uniforms at Civitai
ControlNet -- provides ability to pose a subject based on an outline or other source e.g. Open Poses (predefined free set of poses), Open Poseai (create your own pose in 3D) or more flexible but not completely free at PoseMy.Art (create your own pose in 3D)

CPU Only

One of the big problems is total uncertainty in what the AI is going to create, even based on your prompt and or guided initial images, masks, seed and guide values, for more see a samplers example at ArtStation.

So for the CPU version, I wrote two scripts to help automate the mass creation of images using a local AI install of Stable Diffusion OpenVino which can run on purely CPU (as well as GPU if you have CUDA installed). In this way you can easily create a large number of images and select the ones that most fit your requirements.

See some example images generated on my Dell Precision T3620 (with an NVidia GeForce 1080ti 11GB GPU). To be clear, most of these images were generated with CPU only mode (Vino capability), though the images would be identical to ones generated with GPU, they just took longer to create.

Stable Diffusion Resources

On different machines the install and requirements to run seem to vary, see here for more information on installing.

AI generators in general, and Stable Diffusion specifically, allow you to vary the inital seed, the strength of the noise, an initial image, a mask image, number of iterations, a guidance prompt and strength of guidance of the inital conditions. From testing it is clear that you really cannot be certain what the result will be unless you make a set of images varying the inital values slightly to see the affect. Thus I have written some scripts to assist this, see help (run with -h) for more info on usage.

makeimg.sh - My script to make a single image based on input parameters, this drives demo.py ()
makesetofimgs.sh - My script to make a set of images based on input parameters, this drives makeimg.sh ()
demo_tz.py - My modified version of demo.py to allow creation of multiple images faster than the original ()
batch.csv - An example batch file to make a set of images based on input parameters, this drives makesetofimgs.sh ()
base.set100.csv - An example batch file to make a set of images based on input parameters, this drives makesetofimgs.sh ()
requirements.txt - I had to change the requirements to install on my box, you may have to do the same ()

As you never really know what the AI is going to come up with, try the following for generating a number of images with a different seed for each one

for seed in $(seq 23450 1 23470) ; do ./makeimg.sh -z $seed -s 0.4 -g 8.0 -i 16 -l starship enterprise in orbit over red planet ; done

Here is an

example of the outputs of the above.
or, better for doing multiple images with different control values to see variations
  ./makesetofimg.sh -B -z 1200 -Z 1210 -s 0.4 -S 0.1 -g 4.0 -G 1.5 -i 16 -l starship enterprise in orbit over red planet
or to use an existing set of seeds/variation values
  ./makesetofimg.sh -b mybatchset100.csv -l starship enterprise in orbit over red planet
or to vary only by a seed with the other values the same, try something like
  ./makesetofimgs.sh -B -z 2345 -Z 2355 -s 0.5 -S 1 -g 7.5 -H 7.5 -i 16 -l starship enterprise in orbit over red planet
or a good set of batches to try, varying only the seed and guide as this seems to have the most affect when only a prompt is provided, try

./makesetofimgs.sh -X -B -z 0 -Z 20 -s 0.5 -S 1.0 -g 2.0 -G 3.0 -H 8.0 -i 16 Fairly tale castle with tall white tower next to lake and mountains in background

or a good set of batches to try when using a seed image, with a mask (whole image white with black area of where to keep image) as well as a prompt, try

./makesetofimgs.sh -X -B -z 0 -Z 20 -s 0.8 -S 0.1 -g 3.0 -G 3.0 -H 9.0 -i 16 -I myseedimage.jpg -M mymaskimage.jpg Fairly tale castle with tall white tower next to lake and mountains in background

Here is an

example of the outputs of the "Fairly tale castle..." prompt as shown above.
or if you have a base image, say a photo of yourself, and you want to be in an image try varying seeds and other parameters, this one does 2 strengths and 2 guides using the seeds (0 1000 123450 223450 567000) for the inital image and mask

for ASEED in 0 1000 123450 223450 567000; do BSEED=$(expr $ASEED + 10); echo "Doing with range $ASEED .. $BSEED"; sleep 2; ./makesetofimgs.sh -B -z $ASEED -Z $BSEED -s 0.7 -S 0.2 -g 4.5 -G 6.0 -H 11 -i 16 -I ~/Documents/02/theo02.jpg -M ~/Documents/02/theo02_mask.jpg -l jedi knight fighting drones, photo realistic; done

See my Hints on AI art for hints and help on AI prompts and models

Back
Home