How I Do It - mini-history + currently (Patreon)

Past

I've been using Stable Diffusion and it's like for... about 10 months. My first dabbling was creating a oil-painting style pathway and splicing in / re-rendering lamp posts along that (since dall-e wouldn't do what I wanted) which ultimately turned out nice.

Then used kiri.art to play around with text prompting with different models - pretty awesome and ultimately fed into my first attempts when I got around to setting up A1111 on my new "AI / gaming rig" w/ a 4090

Flows

Text 2 Image

My most basic tool is drafting up text prompts and tweaking it to get at least 1 OK Image - and then sending it off to generate some 40-50 images ... uses to do ~100 but I've reduced that because so often there's gold in the first sets. Now - depending on how I'm feeling - I'll setup hires.fix to do it for all images - or wait to pick out the gold and re-render.

To add variation - I often use one of the wildcard templates to introduce different colors, dress styles, or nationalities.

TIP: nationalities have a great effect of reducing the "sameness" of pictures in some models.

"Prompt magic" or other auto-generation tools for prompts can help, too. Sometimes the nonsense extra terms make things more interesting... and especially help with those big batches of 50 images when I have other things to be doing.

Another tool for prompt generation that I've tried is using a local large language model (think ChatGPT) to give it an idea of what a prompt looks like - and then ask it for some ideas with basic guidance. This had led to some fun images that I recall - need to retry it one of these days with the better models.

T2I / Image 2 Image - ControlNet

This is a very powerful setup with some gotchas. It's a little harder to perform a hires.fix to a certain resolution that I can with just txt2img (vram limitations) but if I have an input that is perfect- I can just use other upscaling techniques.

I've used ControlNet in many ways to great success. Taking images from other places (boorus, morphs, etc) has fun results and works nice as an idea builder - not sure how I feel about "publishing" those. However, games like HoneySelect work great w/ a little know-how to get bodies posed how you'd like and pick clothing/etc.

Now - given an image to base off of - there's a few ways about it:

• Use it with 1 or more of depth-map / normal map / scribble / softedge / open pose to try to get what you want out of it. Really depends.
• Take a processed output (especially scribble or softedge) and send it to Krita to edit (ex: remove details in the lines that are causing problems, add new ones to add new things or modify clothing subtly) and then throw it back in

Image 2 Image / Inpainting

I've sometimes taken model pictures and thrown them through I2I for funsies with nice results. However for my usual case, I'll take an image I generated before or from Honey Select and go through a loop of upscaling / I painting to:

• Fix something that doesn't look right (busted up hands, skin looking wrong, etc)
• Add a new feature - broken straps, clothing tears, potions
• Re-render face to either add more detail or enhance eye color
• Upscale more generally

Sometimes I go about it directly inside Automatic1111's interface- especially if its simple edits. Other times I grab the last round of an image - throw it in Krita - make my edits - paste back into Automatic1111 to continue the process.

Final Bits

Oh - and my final image processing - once I get the image done with in A1111 is to do final fixups:

• Touch-ups for things inpainting wouldn't do. Sometimes manual edit - others the auto-heal tool helps w/ a few tries.
• Crop out problem areas (ex: hands that just wouldn't work - but we're far enough on edge of frame)
• Apply an "unsharp mask" to sharpen the image slightly
• Esp for realistic images: Apply transparent random noise to simulate ISO noise - helps a lot with getting rid of seemingly perfect smooth "plastic" skin