AI Upscaled Textures
Posted: Fri Dec 21, 2018 10:06 pm
UPDATE:
I'm going to try to keep this first post updated with the latest on this project.
---------------------------------------------------------------------
UPDATE:
Satsuki has created a great GUI for ESRGAN. It can even handle pre-filtering! ESRGAN Resizer
Summary:
We are trying to upscale the original Daggerfall textures using a variety of artificial intelligence techniques and some combination of image filters.
All textures have already been upscaled by Freak2121 using Topaz A.I. Gigapixel.
You can download that here: Freak2121's Texture Pack (Gigapixel 4x)
MrFlibble and I have been using various models and pre-processing techniques to enhance upscales using ESRGAN and SFTGAN.
I have created further enhanced textures for Privateer's Hold and created advanced materials for them. You can download this texture replacement mini-mod here (screenshot below):
Masonface's Upscaled Texture Pack (4x + PBR materials) - Privateer's Hold
Note: Make sure filtering mode is set to bilinear or trilinear, not point!
Screenshot: Note: All the textures and more have been rolled into the PVE mod.
---------------------------------------------------------------------
Methodology:
As of now, Freak2121 is getting great results out-of-the-box on textures using Topaz's A.I. Gigapixel.
Other techniques that we've employed are ESRGAN and SFTGAN with mixed results depending on the source image, the model used, and any pre or post processing filters used. These techniques could prove to be the best route if we can train the model on more appropriate data so that it produces results that are higher definition, yet retain the correct aesthetic.
My opinion is that all generated textures/sprites that have been contributed should be made public to the community so anyone can curate their own collection.
The idea there-after is to compile an official collection of textures from all the fore-mentioned techniques, then curate a singular collection that is agreed to best fit for convenience to end-users.
---------------------------------------------------------------------
Original Post:
By suggestion of Luzur, I started tinkering around with some experimental artificial intelligence upscaling methods.
I just started with some basic tests using ESRGAN (https://github.com/xinntao/ESRGAN), and while they aren't perfect, they really ain't half bad! Take a look:
https://imgur.com/zHqCYr0
https://imgur.com/QD83bTw
I think this would be a good way to get a high quality pack of textures rolled out quickly while talented artists take their time remastering the originals (which I think most of us would prefer).
Next is a test using SFTGAN (https://github.com/xinntao/SFTGAN) which I think looks a bit better for textures:
https://imgur.com/YsyXfEe
I'm going to try to keep this first post updated with the latest on this project.
---------------------------------------------------------------------
UPDATE:
Satsuki has created a great GUI for ESRGAN. It can even handle pre-filtering! ESRGAN Resizer
Summary:
We are trying to upscale the original Daggerfall textures using a variety of artificial intelligence techniques and some combination of image filters.
All textures have already been upscaled by Freak2121 using Topaz A.I. Gigapixel.
You can download that here: Freak2121's Texture Pack (Gigapixel 4x)
MrFlibble and I have been using various models and pre-processing techniques to enhance upscales using ESRGAN and SFTGAN.
I have created further enhanced textures for Privateer's Hold and created advanced materials for them. You can download this texture replacement mini-mod here (screenshot below):
Masonface's Upscaled Texture Pack (4x + PBR materials) - Privateer's Hold
Note: Make sure filtering mode is set to bilinear or trilinear, not point!
Screenshot: Note: All the textures and more have been rolled into the PVE mod.
---------------------------------------------------------------------
Methodology:
As of now, Freak2121 is getting great results out-of-the-box on textures using Topaz's A.I. Gigapixel.
Other techniques that we've employed are ESRGAN and SFTGAN with mixed results depending on the source image, the model used, and any pre or post processing filters used. These techniques could prove to be the best route if we can train the model on more appropriate data so that it produces results that are higher definition, yet retain the correct aesthetic.
My opinion is that all generated textures/sprites that have been contributed should be made public to the community so anyone can curate their own collection.
The idea there-after is to compile an official collection of textures from all the fore-mentioned techniques, then curate a singular collection that is agreed to best fit for convenience to end-users.
---------------------------------------------------------------------
Original Post:
By suggestion of Luzur, I started tinkering around with some experimental artificial intelligence upscaling methods.
I just started with some basic tests using ESRGAN (https://github.com/xinntao/ESRGAN), and while they aren't perfect, they really ain't half bad! Take a look:
https://imgur.com/zHqCYr0
https://imgur.com/QD83bTw
I think this would be a good way to get a high quality pack of textures rolled out quickly while talented artists take their time remastering the originals (which I think most of us would prefer).
Next is a test using SFTGAN (https://github.com/xinntao/SFTGAN) which I think looks a bit better for textures:
https://imgur.com/YsyXfEe