A model tried on approximately 20 fractal images for each keyword, with a variety of different styles, it can reproduce an effect similar to a fractal of the corresponding types.

The Menger model is trained on a variety of 3D renders, while the Sierpinski uses a mix of 2D and 3D images. For some reason, they tend to produce similar outputs sometimes.

For the Menger images I tried this simple prompt:

spectacular mengersponge castle entrance view, 4k trending detailed render, volumetric lighting, cinematic octane render

For the Sierpinski images I tried this simple prompt:

spiked ornate triangle abstract art, sierpsponge, colorful octane render, realistic 4k