The Gaming Restoration | RESIDENT EVIL 3 Intro / Opening (Remastered in 4K using AI Machine Learning) @TheGamingRestoration | Uploaded 5 years ago | Updated 5 minutes ago
Remastered in 4K using Machine Learning
I took the original video files from the Biohazard 3 SOURCENEXT PC port, and used ESRGAN with the Manga109
training model to upscale the videos 4 times their original resolutions…
This time the results were mostly disappointing.
I'm not sure if it's because of the decoding software that Capcom used back in the 90s, but RE3 FMVs look
incredibly "blocky" and pixelated, and ESRGAN was unable to reconstruct the frames properly (it was able
to sharpen the edges but it couldn’t enhance the details). To make matters worse, the videos
are interlaced, and all the characters and objects have a great deal of motion blur.
On the bright side, I tried using STRGAN (another neural network software that can recover lost details in
textures and pictures) and I was pleased to see some beautiful results. It didn't fix any of the issues with
the videos, but it was able to magically enhance some details like clothing, the zombies faces and their
skin details, the asphalt on the streets, the blood stains... basically any grungy-looking material/texture.
These results can be better appreciated in picture than in motion though, so I will start
this video by showing a small comparison, and then the opening FMV.
Enjoy.
Original resolution: 720x480
[Deinterlaced and downscaled to 640x480 to remove excessive blocking and aliasing]
ESRGAN Upscale: 2560x1920
0:20 SFTGAN&ESRGAN vs original
1:19 Opening FMV (Upscaled)
ESRGAN: github.com/xinntao/ESRGAN
SFTGAN: github.com/xinntao/SFTGAN
Manga109Attempt.pth: http://www.mediafire.com/file/w3jujtm752hvdj1/Manga109Attempt.pth.zip/file
Remastered in 4K using Machine Learning
I took the original video files from the Biohazard 3 SOURCENEXT PC port, and used ESRGAN with the Manga109
training model to upscale the videos 4 times their original resolutions…
This time the results were mostly disappointing.
I'm not sure if it's because of the decoding software that Capcom used back in the 90s, but RE3 FMVs look
incredibly "blocky" and pixelated, and ESRGAN was unable to reconstruct the frames properly (it was able
to sharpen the edges but it couldn’t enhance the details). To make matters worse, the videos
are interlaced, and all the characters and objects have a great deal of motion blur.
On the bright side, I tried using STRGAN (another neural network software that can recover lost details in
textures and pictures) and I was pleased to see some beautiful results. It didn't fix any of the issues with
the videos, but it was able to magically enhance some details like clothing, the zombies faces and their
skin details, the asphalt on the streets, the blood stains... basically any grungy-looking material/texture.
These results can be better appreciated in picture than in motion though, so I will start
this video by showing a small comparison, and then the opening FMV.
Enjoy.
Original resolution: 720x480
[Deinterlaced and downscaled to 640x480 to remove excessive blocking and aliasing]
ESRGAN Upscale: 2560x1920
0:20 SFTGAN&ESRGAN vs original
1:19 Opening FMV (Upscaled)
ESRGAN: github.com/xinntao/ESRGAN
SFTGAN: github.com/xinntao/SFTGAN
Manga109Attempt.pth: http://www.mediafire.com/file/w3jujtm752hvdj1/Manga109Attempt.pth.zip/file
![Cyberpunk 2077 - Teaser Trailer (2013) (Remastered in 4K using AI Machine Learning) [IMPRESSIVE]](https://i.ytimg.com/vi/LBjBTXjgN88/hqdefault.jpg "Cyberpunk 2077 - Teaser Trailer (2013) (Remastered in 4K using AI Machine Learning) [IMPRESSIVE]
Cyberpunk 2077 - Teaser Trailer (2013)
Remastered in 4K using ESRGAN [AI Machine Learning] [IMPRESSIVE RESULTS]
ESRGAN is an AI Neural Network software that is capable of upscaling and reconstructing low-resolution images, while retaining most of their elements and details intact.
I took the original uncompressed trailer that CD Projekt released back in 2013, and used ESRGAN with the Manga109 model to upscale the video 2 times its original resolution.
While the software is perfectly capable of upscaling 1080p images to 8K [7680x3200], the current models werent trained for such high resolutions.
Once the image reaches a certain amount of pixels, the AI cant improve the details any further.
Original resolution: 1920x800
ESRGAN Upscale: 3840x1600
Logo made by crepet9000 (DeviantArt): https://www.deviantart.com/crepet9000/art/Cyberpunk-2077-Logo-764106006")
![Fatal Frame - 4K 60fps Gameplay [Graphical fixes] (PCSX2 1.5.0)](https://i.ytimg.com/vi/LhF9rTObxQo/hqdefault.jpg "Fatal Frame - 4K 60fps Gameplay [Graphical fixes] (PCSX2 1.5.0)
PCSX2 1.5.0-20190415172425 - compiled on Apr 23 2019
OS: Linux 5.0.8-20-tkg-pds
Physical RAM: 6GB
CPU: Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz
GPU: Nvidia GT 1030 2GB
NOT RECORDED IN REAL TIME
GSdX OpenGL (Hardware) Rendered at 6x Native Internal Resolution (3840x3072)
MTVU + Fxaa shader + No deintarlacing hack + Widescreen hack (plus fixes)
PCSX2 Config
MTVU On
GSdx
OpenGL (Hardware)
Internal Resolution 6x Native
Automatic CRC Level Partial (OpenGL Recommended)
Blending Unit Accuracy High (Recommended high-end PC)
Shader Configuration ) Enable FXAA
60FPS Hack code:
339A0B8C.pnach
gametitle=Fatal Frame (SLUS-20388) 0x339A0B8C
//Disable interlacing
patch=1,EE,2023188C,word,00000000
**FILTERS & WIDESCREEN HACK CODES ARE ALREADY INCLUDED IN THE PCSX2 EMULATOR")
![Resident Evil: Dead Aim - 4K 60fps Gameplay [Graphical fixes] (PCSX2 1.5.0)](https://i.ytimg.com/vi/RxZkksmJgjw/hqdefault.jpg "Resident Evil: Dead Aim - 4K 60fps Gameplay [Graphical fixes] (PCSX2 1.5.0)
·Align Sprite + Merge Sprite (Fix vertical lines)
·Auto flush + Vertex Half-pixel Offset (Fix bad post-processing)
·No interlacing code [by asasega]
·Widescreen hack
PCSX2 1.5.0-20190312131301 - compiled on Mar 14 2019
OS: Linux 5.0.1-3-tkg-pds
Physical RAM: 6GB
CPU: Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz
GPU: Nvidia GT 1030 2GB (GDDR5)
·GsdX OpenGL (Hardware) – Rendered at 6x Native Internal Resolution (3840x2874)
·Custom External shader conf
FBB5290C.pnach
gametitle=Resident Evil: Dead Aim [SLUS-20669] (U)
//Disable interlacing
patch=1,EE,2028A268,extended,00000050
patch=1,EE,2028A274,extended,000001E0
patch=1,EE,2028A284,extended,00000000")
![Silent Hill 2 - A Francis Bacon Painting [Using AI Machine Learning to get creepy results]](https://i.ytimg.com/vi/SvRNh39RYBY/hqdefault.jpg "Silent Hill 2 - A Francis Bacon Painting [Using AI Machine Learning to get creepy results]
Silent Hill 2 - A Francis Bacon Painting
Using ESRGAN to fully reconstruct videos [AI Machine Learning]
ESRGAN is an AI Neural Network software that is capable of upscaling and reconstructinglow-resolution images, while retaining most of their elements and details intact.
This is an experiment I did to show what happens when an AI tries to reconstruct a video with almost no pixels or information in it.
Basically, I took some of the FMVs from the original Silent Hill 2 disk, and downscaled them by ~80%. Then I grabbed a training model that has dozens of references of human faces, and used ESRGAN to upscale the footage 8 times its original resolution.
The results are pretty abstract and disturbing, and they almost resemble the aesthetics of Francis Bacon (Masahiro Itos favorite painter, whose works served as inspiration for the monsters designs)
Original resolution: 512x384 → 64x48
ESRGAN Upscale: 1024x768 → 1440x1080
(Re-upscaled using Lanczos and temporal sharpening)")
