markec

Free Play Account
  • Content count

    44
  • Joined

  • Last visited

Everything posted by markec

  1. Found an interesting graph for the training. Seem i'll just train it to 100k as from 100 to 500k there is only 0.5 PSNR difference which is pretty minimal (the detail accuracy ratio), as the program starts from 20 PSNR i think, so from 0 to 100 it's 10 PSNR and from 100 to 500 only 0.6 lol.
  2. Hello, i'm a long time lurker, I registered to make everyone aware and especially the developers of an open source upscaler called ESRGAN (Enhanced Super-Resolution Generative Adversarial Networks). What it is? Basicaly it's 'zoom and enhance'. It can be used to greatly improve low quality textures. I don't have access to direct textures, so I took a screenshot of a texture in-game, and then enhanced it to 4k with ESRGAN, the difference is obvious. Imagine if I can achieve such result with just a screenshot, what could the devs achieve with their original textures. Here is the helmet texture I used After ESRGAN Full enhancement What the devs would need to do? All they have to do is run all of their textures through ESRGAN (or SFTGAN which is a different version) and then put them back into the game. links: ESRGAN - https://github.com/xinntao/ESRGAN SFTGAN - https://github.com/xinntao/SFTGAN tutorial on how to use ESRGAN/SFTGAN - https://kingdomakrillic.tumblr.com/post/178254875891/i-figured-out-how-to-get-esrgan-and-sftgan
  3. It's automated, with every new iteration the system makes more accurate interpretation based on the source material. Basicaly it's like real life learning.... the more you do something, the better you get at it
  4. After training it for 4 hours on 50 textures that I have as a test, i can already see the potential, when I have a big enough texture set and train it fully, it should be giving perfect representation of low res textures.
  5. So i've learned how to 'train' the program, need around 100 textures to train it to recognize the ww2ol texture patterns (the more the better), the actual training takes time though, i've read that the authors trained esrgan for 500,000 iterations, my PC processes 200 iterations in 7 minutes, so it will take me 2-3 weeks of real time to 'train' the AI and the result which will be produced is unknown, but it's worth a try
  6. In the mean time while waiting for ground textures, i took some from a different game for a test and ground textures look great upscaled
  7. Latest batch is done, eliminated over processing (nr 2) completely by using a different upscaling model, now waiting for Hatch to do the magic
  8. I'm doing now a 3rd pass on the textures because i wanted to get more sharpness and detail and this is the effect
  9. That's because both are 2k for comparison, the original texture is 256x256 and it looks so pixelated at 2k. The same thing happens in the game when you look at an object with low res texture, it looks blurred out without detail...... when you will see in game you will see the difference.
  10. It was the borg ! in other news, i've used a different method on the textures which provides a better accuracy, Hatch should already have the improved textures.
  11. Ok so i've done all the textures Hatch sent me and we should see some before/after shots from him
  12. So i have noticed that the flag had problems because it was very low res 128x128, the upscaler struggles with anythin at 128x128 and below to maintain the details, at 256x256 and above it's good.