Handprints Spanking Art & Stories Page Drawings Gallery #285

Spankart writes:

I start with an idea for an image, a concept. Early on, I also decide on a style, such as oil painting, watercolor painting, or drawing. I always make sure to steer the AI away from photorealism because that's a no-go for me; most of my images feature a minor spankee, so it has to be a digital drawing or paintingï¿½art that is stylized, somewhat realistic and somewhat abstract, definitely not photorealistic.

I use an AI of my choice to generate starting images, refining my prompt until I get something good enough for me. Any AI will do. I usually make my starting images SFW, so censorship is not usually an issue. So far, I have used Stable Diffusion, Reve, FLUX.1, Qwen-Image, and one or two others. New AI models are coming up all the time, so I just use whatever I find does the job well. I rarely use LoRA. They don't work well with my artistic concepts and workflow; I don't need them.

Then comes the main work, which usually keeps me busy for a week or two. First, I manually edit the image using Gimp. I do whatever I want to change; I have become good with Gimp over the years. Next, I export the image and use it as input in Stable Diffusion's img2img inpainting mode. I have ForgeUI with SD 1.5 installed locally, and I use a model that's good for inpainting, as well as for the styles and content of my artwork. My currently preferred model is named cyberrealistic_v41-inpainting.

I select an area to inpaint, a prompt for the inpainting job, and a denoising strength between 0.4 and 1, depending on the job. I always run a batch of 6 images. Sometimes I use ControlNet OpenPose or NormalMap to force the inpainting to remain tighter to what's already in the image. Then I select between 1 and all 6 of the output images and add them to my working image in Gimp, as new layers on top. I select the topmost layer and use the eraser tool with a blurry brush to erase everything that's not best in this topmost layer. When this is done, I merge down to reduce the number of layers, and repeat. Now I have a slightly improved version of my image. Again I do manual editing if I find that helpful. Then comes the next export, the next round of inpainting, and bringing the improved results into the working version.

In 2022, I made this tutorial about my workflow: https://spankingart.org/wiki/File:Spanking-art-with-stable-diffusion-tutorial-by-spankart.png This tutorial is somewhat outdated. Today I work in a higher resolution from the start (e.g. 1024 ï¿½ 1536 pixels), and don't upscale anything near the end. But I still use the old-fashioned SD for the inpainting although better AIs are available now.

Here's a time-lapse animation showing the making of one of my images in 2024: https://spankingart.org/wiki/File:Bared_WIP_anim.mp4

Adding a character to an image that blends seamlessly into the rest is not easy, but it can be done. Fixing an arm, leg, hand, or face if the rest is good is relatively simple. Removing clothing is easy, too. It's easier than adding pulled-down pants, underwear, and other uncommon things, but these things can be done too, so long as you guide the AI firmly enough with pixels, prompts, and relatively low denoising strengths. Unfortunately, my model of choice struggles immensely with soles of feet. I still need to find a better model for that.

Although I label all of my artwork as AI-generated, as you can see from this workflow description, it involves a lot of manual work. Every artwork is the result of many (5-20) hours of work. For the most part, I only use the AI to start the artwork and refine the details to achieve better image quality than I could manually. I'm happy to teach my methods to anyone. Just contact me on Discord (@Spankart). However, I can't teach you an easy, fast method that doesn't involve a lot of manual work. AI tools are great, but if you want more than simple text-to-image generation, you still need to put in time and effort. In return, though, you get to really control the output.