There's definitely improvements that can be done. Those cloudy kind of patterns and other visual glitches are generally from using training images with compression artifacts. So it's best to use as high quality of images as possible, even if you have to downsize them by a lot. Another part of the issue is the downsizing process itself. Most image downsizers that make the images the appropriate size for training, use a cheap method that results in a blurrier image (bilinear), so fine details are lost, and the patterns of that ends up in the training too. The thing about training is it will consider every artifact as part of the image style and reproduce it. For this style I would use "Nearest neighbor (hard edges)" or "Bicubic sharper" for downscaling images to the appropriate training size, and I would also try to find as high of quality images as possible, avoiding JPG format entirely if possible.

For anyone wondering why you cant get access to Triggers and Metadata in Automatic1111 with this Lora. Remove the ' in 1970's safetensors filename and it works again. Otherwise give it a try. Its an exiting LORA :-)