Intro
Another short article in advanced series. This time around we'll be applying underpainting as foreground enhancement to an already formed composition.
If you've been generating AI images for a while, you probably already noticed, that compositions involving characters are quite static and rarely have anything obscuring the main subject. This is because diffusion model mimics the training data, which almost 100% of the time has subject in clear view and unobscured. Doesn't mean it can't do foreground objects. In this article we will explore how to add some "magic" and "splash" to your image by adding particles to the foreground.
It's not very complicated and in essence is just scrible inpaint with extra steps. But we'll do it with less control and in the spirit of img2img randomness.
Let's go
Step 1 - Generate a composition. You can use txt2img or img2img - anything you like. Just don't upscale it just yet.
Step 2 - Drop the image into Krita.
Step 3 - Paint your overpainting.
- This part can be very simple. You can just create a new layer, pick a texture brush with a contrasting color (compared to the image) and doodle randomly all around. Sometimes that's enough for something basic, like a snowfall, for example.
- Or... You can setup alpha layer group with a colorful underpainting just like in Vignette article. And reveal some of that underpainting with a texture brush. You want high saturation and brightness on the underpainting in this step. Underpainting will add gradient colors and texture.
(I had falling leaves in mind, in this one)
(Something colorful)
Protip1: Doesn't matter if you don't know what you want to appear in the foreground, let Stable Diffusion figure it out for you. But if you do have something in mind (like lighting, foliage, petals, snow or rain), then make sure your doodle color and shape is vaguely reminiscent of it.
Step 4 - Now that you have your image with overpainting on top, save it and go back to Stable Diffusion img2img tab.
- From here, you can either denoise entire image at low-medium strength.
- Or... If your composition has parts you really want to save (face, hands, feet - anything) - mask them in inpaint, and denoise only unmasked part of the image instead.
Protip3: If you had some foreground objects in mind - add keywords to the front of the prompt.
Protip4: Don't be upset if it doesn't work out right away. It's very hard to predict for a human how colors and shapes are going to interact in an AI model. Switch back to Krita and adjust noise color, delete some, add some, blur or distort the shapes. You can even do multiple layers of overpainting.
Step 5 - Give it a few batches, just like underpainting, it takes a few tries to find the perfect one.
Step 6 - Ultra upscale.
???
Step 7 - Profit.
Protips
Protip5: During creative part, try to use less Negative Embeddings. Embeddings will make "too much sense" of our noise. Meaning, stacked negative prompt will almost always turn green noise into leaves, for example.
Protip6: Denoise at medium strength (0.45+) to blend noise into your image entirely. This can cause various effects, most often recoloring or extra detail.
The End
This method works for anything you can think of and works even better for formless/abstract shapes like energy currents and magic for fantasy images. There're probably many more applications for partially covering image with an overpainting, but you'll have to discover that on your own.
That's it, you reached the end of the article.
All images done in this article were made with UltronMerge (male-centric/homoerotic model).
Good luck and have fun!