Anime4K
Anime4K copied to clipboard
bloc97
Improve performance for stable v1.0 release.
Improve line detection speed as it does not need to be pixel precise. The underlying line refinement and FXAA algorithms are already robust enough. False positives are less harmful than false negatives. (eg. use smaller gaussian kernel, or even use lower luma resolution)
Use line detection to skip processing some parts of the picture all together when the probability of finding a line is lower than some threshold. (eg. 10% when refining and applying FXAA)
Combine similar passes into a single pass if possible. This will reduce the time needed to copy textures between separate passes. On 4K+ resolutions image processing is often memory-bound and not compute-bound.
The target should be 4-5ms.
I am Japanese. I can't write English. (Machine translation) Thank you for the wonderful software. I would like to ask a few questions. Please give me an answer if you like.
1.Are there any plans to develop parameters optimized for [P720 to P1080] and [SD480 to P1080]? 2. Are there any plans to implement parameters that can be adjusted by the user? 3. Is there a plan to conduct a wide range of user tests by releasing multiple parameters?
This software is light, clean and very easy to use I can't write a program. As a tester, you may be able to report the results.
Have a nice day!
@PULU38 1- Yes, we do plan on optimizing lower resolutions. However, we think machine learning would be necessary since we would need to recover texture detail. We plan on using a hybrid approach where ML will be used to optimize this algorithm's hand crafted parameters.
2- You can already change the parameters in the code, but they are not well documented. I think we should be able to write a small guide for those who want to tweak the settings.
3- I don't think we have enough time and resources to perform large scale testing.
Thank you for the questions and suggestions.
i have tried simple structure cnn(with resblock) using edge guidance,which looks good on optimizing lower resolutions.
