Stable Diffusion with Automatic1111

I started playing around with the WebUI, trying different combinations of keywords, or "prompts" as they're called. Really, I was just tinkering to learn the UI, but I hadn't really done anything that meant anything, just a bunch of random stuff to see what would happen. Like a friend telling me about the game "Stray" in a conversation, so I would put together a prompt like "vaporware stray cat in a cyberpunk dystopian alley" and creating this cool 4 image bundle:

I was feeling the itch. By this point, I already knew I was hooked.

But now I wanted to do something a bit more involved than just throwing random words together and seeing what the system spits out. I needed to have a specific direction I wanted to go.

I had recently downloaded the mobile time killing paint-by-numbers app, Dark Colors, and had saved one of the pieces of artwork of D.C. comic villainess, Harley Quinn:

I had this image on my phone and wanted to try a few things out anyway, so I connected to my network and used the Img2Img feature, making sure to set the "Denoising Strength" all the way down to 0 (zero) so I could see just how exact it matches. In other words, is this still AI generated or is it simply an image copy. 

It has some minor flaws, but I knew this was a real deal A.I. generated image, and not just a photocopy of the original.

Looks pretty damn good, doesn't it? You can notice things like the eyes or the letters in the watermark, even the fishnets are a bit washed out. But this was a great test! Also, it put the image in my render history, so now I could just pull the image from there using the "Send to img2img" feature on the "History" tab of my WebUI instance.

The high denoising value gets a closer background to what I was wanting, but Harley just feels out of place here.

And the low denoising value has all kinds of other problems, mostly that while Harley looks closer to the original image, she has problems with what she's holding, the changes in her clothes, etc. But more than that, the background is awful.

I mostly hated how dark and red the backgrounds kept ending up.

This is where we get into another feature: Inpaint

Inpaint gives you the capability to draw a mask on top of your image before using that image for processing. You also have a few cool options with Inpaint that allow you to either:

1. Only inpaint everything that has been masked, or

2. Only inpaint everything that has not been masked.

It was easier for me to trace over the figure of Harley and go with option 2:

Now that I had Harley masked, I decided to crank the Denoising strength all the way up. Also, since I was effectively removing Harley from the new render, I removed her from the prompt, making the key subject material be "standing in baseball park."

My full settings were as follows:

• Prompt: Masterpiece, best quality, standing in baseball park

• Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry

• Steps: 20, Sampler: Euler a, CFG scale: 12, Seed: 3445035576, Face restoration: GFPGAN, Size: 512x512, Model hash: 587f05f3, Denoising strength: 1, Mask blur: 4

And this gave me something I was ready to work with!

Hell yeah! We have Harley standing on the field! And she looks good!

Good... but not "great." And I'm not cool with my girl Harley only being "good" so I gotta do something about that...

As you can see, the background was a success, but I missed or didn't adequately cover areas with my mask, so she's a bit rough around the edges.

Also, I wanted to push this. I wanted to see this flat image as a 3D styled render with some realism.

So I added that to my prompt. And since we now wanted the AI to recognize Harley for who she is, I also put her name in the prompt.

The complete settings were:

• Prompt:Masterpiece, best quality, harley_quinn satisfied standing at ballpark 3d realism

• Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry

• Steps: 20, Sampler: Euler a, CFG scale: 12, Seed: 3679158294, Face restoration: GFPGAN, Size: 512x512, Model hash: 587f05f3, Denoising strength: 0.5, Mask blur: 4

And the results... just wow!

These were good...

Again, not great... good... but what wasn't right about them?

The problem with these was the subject size relative to the canvas. Because Harley takes up so much of the frame, it leaves very little room for anything to fit well. So I decided I needed to change my approach on the theme.

I thought, "how can I frame her in a way that sticks to the crime-in-progress theme?"

Grand Theft Auto

So now we're not just looking at a picture as an observer of art... we're engaged in the art! We're about to get carjacked!

New settings:

• Prompt: Masterpiece, best quality, POV looking into car

• Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry

• Steps: 20, Sampler: Euler a, CFG scale: 12, Seed: 3445035576, Face restoration: GFPGAN, Size: 512x512, Model hash: 587f05f3, Denoising strength: 1, Mask blur: 4

Result:

Well... shit... I didn't even think that POV looking into car would be interpreted as "from my point of view as if I'm looking into a car"

Quick fix on the prompt: Masterpiece, best quality, POV looking out from car

This didn't quite look right, so I modified the prompt just a little bit more: 

Masterpiece, best quality, POV looking out from car window