"Fullscreen Image Enhancement Pack" Discussion

@AeroWidescreen @elishacloud

Holy cow guys, I needed to share this with you right away because I'm blown away by the results. I happened to stumble across an image enhancement site called "Let's Enhance." I'll let their site explain it best:

Let’s Enhance uses cutting-edge Image Super Resolution technology based on Deep Convolutional Neural Networks. Before appearance of this technology it was impossible to dramatically increase photo or image size without losing quality. Your best option in Photoshop, called Bicubic Interpolation - made your image unsharp and blurry. Those of you on the math side of the things could argue – however you increase image resolution – there is no new information about the image - you just can’t add extra quality! This is not true in a case when Neural Network and AI is used. Our neural network is trained on a huge dataset of images, so it learns typical features of physical objects – bricks of walls, hair and skin. After the network recognises those features on the images uploaded and add extra details, based on its general knowledge the the world.

So... I ran a few test images through it and... wow. I think we struck pay dirt here, friends. Check out the images below to see what I mean (be sure to click them to view them at their true sizes). I used the software Texmod to extract the PCSX2 (PS2) version of these fullscreen images to guarantee I'm using the best version of these images.

It seems to struggle a little on small lettering but, can you blame it? I honestly can't believe it's able to read some of those words at all from the original, low res versions. But those parts can be fixed up some in Photoshop after the fact.

I don't know about you guys but I'm impressed. If it's possible to pursue a "Fullscreen Enhancement Pack" mod down the line I'll pay for a package/membership with this site so we can run the images through it.

Default image (extracted from PCSX2):

Resampled (4x) through Photoshop (via Bicubic (smooth gradients)):

Processed/resampled (4x) through "Let's Enhance":

elishacloud / Silent-Hill-2-Enhancements