Difference between revisions of "Different Neural Networks"

From Upscale Wiki
Jump to navigation Jump to search
(Simple port of the Github wiki page.)
 
m
Line 1: Line 1:
==Different Neural Networks==
 
 
 
There are a lot of different neural networks, in fact there are new ones every day. But here is a ''small'' list of important ones:
 
There are a lot of different neural networks, in fact there are new ones every day. But here is a ''small'' list of important ones:
  
===Upscalers===
+
==Upscalers==
 
----
 
----
====ESRGAN (free)====
+
===ESRGAN (free)===
 
[https://github.com/xinntao/ESRGAN ESRGAN] is currently one of the best tools to upscale any texture. There are multiple models beside the default one, if you want to try one of those look in <code>#esrgan-model-links</code> on our [https://discord.gg/SxvYsgE Discord server]. With a custom model it can outperform more or less any other tool out there currently.
 
[https://github.com/xinntao/ESRGAN ESRGAN] is currently one of the best tools to upscale any texture. There are multiple models beside the default one, if you want to try one of those look in <code>#esrgan-model-links</code> on our [https://discord.gg/SxvYsgE Discord server]. With a custom model it can outperform more or less any other tool out there currently.
  
Line 12: Line 10:
 
Note: To prepare Datasets for training [https://github.com/deorder/texture-upscaler Deorder's texture-upscale scripts] are recommended.
 
Note: To prepare Datasets for training [https://github.com/deorder/texture-upscaler Deorder's texture-upscale scripts] are recommended.
  
====SFTGAN (free)====
+
===SFTGAN (free)===
 
[https://github.com/xinntao/SFTGAN SFTGAN] is from the same developers as ESRGAN. It works by separating an image into different sections before it upscales. Feel free to try it.
 
[https://github.com/xinntao/SFTGAN SFTGAN] is from the same developers as ESRGAN. It works by separating an image into different sections before it upscales. Feel free to try it.
  
Line 19: Line 17:
 
'''Note:''' There might be a few things missing for segmentation training.
 
'''Note:''' There might be a few things missing for segmentation training.
  
====waifu2x (free)====
+
===waifu2x (free)===
 
[https://github.com/nagadomi/waifu2x waifu2x], [https://github.com/lltcggie/waifu2x-caffe with GUI] isn't as good in my opinion, ESRGAN and A.I. Gigapixel both upscale better than all models from waifu2x in my experience. It also noise and jpg artifact cleaning.
 
[https://github.com/nagadomi/waifu2x waifu2x], [https://github.com/lltcggie/waifu2x-caffe with GUI] isn't as good in my opinion, ESRGAN and A.I. Gigapixel both upscale better than all models from waifu2x in my experience. It also noise and jpg artifact cleaning.
  
====Gigapixel AI (non free)====
+
===Gigapixel AI (non free)===
 
[https://topazlabs.com/gigapixel-ai/ Gigapixel AI] formerly A.I. Gigapixel is a commercial product from topazlabs and unlike ESRGAN or waifu2x costs money. It doesn't reach ESRGAN's results in upscaling if you use a custom model, but it works quite well on photos and skin and is easy to use. It was made with photos and not textures in mind. It tends to be somewhat grainy, but allows control over deblurring and denoising.
 
[https://topazlabs.com/gigapixel-ai/ Gigapixel AI] formerly A.I. Gigapixel is a commercial product from topazlabs and unlike ESRGAN or waifu2x costs money. It doesn't reach ESRGAN's results in upscaling if you use a custom model, but it works quite well on photos and skin and is easy to use. It was made with photos and not textures in mind. It tends to be somewhat grainy, but allows control over deblurring and denoising.
  
===Image artifact Cleaning and improvement===
+
==Image artifact Cleaning and improvement==
 
----
 
----
====ESRGAN again (free)====
+
===ESRGAN again (free)===
 
[https://github.com/xinntao/ESRGAN ESRGAN] again is currently one of the best tools to clean any texture. There are multiple models for that, if you want to try one of those look in <code>#esrgan-model-links</code> on our Discord server.
 
[https://github.com/xinntao/ESRGAN ESRGAN] again is currently one of the best tools to clean any texture. There are multiple models for that, if you want to try one of those look in <code>#esrgan-model-links</code> on our Discord server.
  
Line 34: Line 32:
 
Note: To prepare Datasets for training [https://github.com/deorder/texture-upscaler Deorder's texture-upscale scripts] are recommended. The <code>train.py</code> must also be slightly modified to train 1x models.
 
Note: To prepare Datasets for training [https://github.com/deorder/texture-upscaler Deorder's texture-upscale scripts] are recommended. The <code>train.py</code> must also be slightly modified to train 1x models.
  
====Sharpen AI (non free)====
+
===Sharpen AI (non free)===
 
[https://topazlabs.com/sharpen-ai/ Sharpen AI] is another tool from topazlabs, there is also <code>DeNoise AI</code> and <code>JPG to RAW AI</code>, but the later two don't work well for textures in my experience. I am not the biggest fan of Sharpen AI, but a lot of people like it.
 
[https://topazlabs.com/sharpen-ai/ Sharpen AI] is another tool from topazlabs, there is also <code>DeNoise AI</code> and <code>JPG to RAW AI</code>, but the later two don't work well for textures in my experience. I am not the biggest fan of Sharpen AI, but a lot of people like it.

Revision as of 15:53, 12 August 2019

There are a lot of different neural networks, in fact there are new ones every day. But here is a small list of important ones:

Upscalers


ESRGAN (free)

ESRGAN is currently one of the best tools to upscale any texture. There are multiple models beside the default one, if you want to try one of those look in #esrgan-model-links on our Discord server. With a custom model it can outperform more or less any other tool out there currently.

If you want to train ESRGAN head here. It also contains code for testing.

Note: To prepare Datasets for training Deorder's texture-upscale scripts are recommended.

SFTGAN (free)

SFTGAN is from the same developers as ESRGAN. It works by separating an image into different sections before it upscales. Feel free to try it.

If you want to train SFTGAN head here. It also contains code for testing.

Note: There might be a few things missing for segmentation training.

waifu2x (free)

waifu2x, with GUI isn't as good in my opinion, ESRGAN and A.I. Gigapixel both upscale better than all models from waifu2x in my experience. It also noise and jpg artifact cleaning.

Gigapixel AI (non free)

Gigapixel AI formerly A.I. Gigapixel is a commercial product from topazlabs and unlike ESRGAN or waifu2x costs money. It doesn't reach ESRGAN's results in upscaling if you use a custom model, but it works quite well on photos and skin and is easy to use. It was made with photos and not textures in mind. It tends to be somewhat grainy, but allows control over deblurring and denoising.

Image artifact Cleaning and improvement


ESRGAN again (free)

ESRGAN again is currently one of the best tools to clean any texture. There are multiple models for that, if you want to try one of those look in #esrgan-model-links on our Discord server.

If you want to train ESRGAN head here. It also contains code for testing.

Note: To prepare Datasets for training Deorder's texture-upscale scripts are recommended. The train.py must also be slightly modified to train 1x models.

Sharpen AI (non free)

Sharpen AI is another tool from topazlabs, there is also DeNoise AI and JPG to RAW AI, but the later two don't work well for textures in my experience. I am not the biggest fan of Sharpen AI, but a lot of people like it.