[oceanofsolaris]

The underlying problem these elaborate prompts seem to solve is that the internet contains many pictures, few of which look very beautiful.

If you look at all internet pictures of sheep, many of them will not be very exciting and depict a low contrast sheep in a foggy landscape.

So to get a picture with strong saturation and clear lines, it helps to put text there that is usually associated with pictures that have these ... like "HD wallpaper" or "made with unreal engine". Most "wallpapers" might be of dubious artistic quality, but muted colors and a lack of saturation will generally not be their problem.

This is of course not the only problem with the model. It doesn't even produce a clear image of a sheep .... but that will probably get better with larger models and more training. Similarly it doesn't seem to have a sense of overall composition and tends towards fractal or tiling-like images. But those problems are probably orthogonal to the fact that the model doesn't per se try to make good pictures ... just average ones for the description you give it.