Midjourney just release their latest Upscaler for v5 images that can take your standard (1:1) 1024px image to 2x and 4x resolution while maintaining the details and creating artifacts. Well that’s the concept so one needs to explore this and see what is the big difference and it it really good?
The PNG files produced a certainly much larger. Now to ensure your retrieve the full size PNG and not the one Discord will show you make you click that Open in Browser link that will load the full size PNG which you can download.
For the sample image that I analysed the following were the file sizes, certainly these will be taking a lot more space on your computer or tablet than before.
For a visual reference this is the size different of the 3 different sizes produced by Midjourney of the same image. The inner most is the standard 1x (1024px wide) image, while the next one is twice as big 2x (2048px wide) and finally the encompassing image is 4x (4096px wide) along the edge.
In my review I found the 4x to introduce lot of snake like artifacts in the image and is a lot less crisper than the 2x image. This is quite visible in skin texture and eye lashes when you zoom into 1:1 ratio of the image. However if you are not looking at the 1:1 ratio zoom of the image then I would say that 4x image is good enough for screen usage, displaying on a website etc.
I have not yet printed these images to see if this makes a difference but I have certainly tried to take the 2x image in Topaz Gigapixel which does a much more decent job at preserving the details when taking the 2x image to 4x (increasing it by 2x).
However this is not about Gigapixel and I was feeding my curiosity. It is about the images that Midjourney produces with these new upscalers.
Personally I prefer the 2x size myself and I’d use external image upscaler in Automatic1111 or Gigapixel to upscale the image. However for most people I think the 4x which is slightly softer will go unnoticed you don’t have to worry to much. I think as I’ve been working with photographs and images for over a decade I generally lean towards the quality over quantity which in this case translates to 2x sized image vs 4x.