Results 1 to 23 of 23

  1. Post
    #1

    Soul Reaver, Final Fantasy VIII (and others) intros remastered in 4K

    If you've not been following over the past months we've seen a lot of texture mods coming out for older games thanks to ESRGAN AI and other machine learning techiques. I saw these today and thought them well worth a share. Some ancient intros remastered using AI. Impressive stuff.

    This channel https://www.youtube.com/channel/UC33...UZFAkMcCCwjyWg






    Some of the texture mods a pretty ****ing amazing too. Remember these are almost entirely done using AI with only some clean-up required by the modder for some noise or glitchiness. Making it a much less labour intensive task and far superior to upscaling like some older texture mods.

    Quake 2
    Click image for larger version. 

Name:	HmJ2pFa.jpg 
Views:	41 
Size:	248.8 KB 
ID:	225357

    Resident Evil 3
    Click image for larger version. 

Name:	xUgj2Y0.jpg 
Views:	40 
Size:	874.0 KB 
ID:	225358

    Max Payne
    Click image for larger version. 

Name:	kaZv2yz.jpg 
Views:	45 
Size:	402.4 KB 
ID:	225359Click image for larger version. 

Name:	FlntVDj.jpg 
Views:	44 
Size:	573.1 KB 
ID:	225360

  2. Post
    #2
    Now, if Soul Reaver could be remastered, that'd be great.

  3. Post
    #3
    hmmm is there any more info from this author? No info on the channel or vids

  4. Post
    #4
    /edit, different channel so posting it






    Not really the right thread for it, pretty sure it has been posted before but man looks so good.




    Onto the real subject, wish SE would take this in update the ports of FFVII/IX with the better backgrounds. Sad that it'll be stuck to the PC versions via modding, at least FFVIII will be done though since SE won't re-release that game.
    Last edited by XSpike; 5th March 2019 at 6:21 am.

  5. Post
    #5
    Very interesting to see the differences between the ML techniques as well. Check this comparison from FFIX out, for example. Megapixels looks closer to the original, even closer to ESRGAN trained with FFIX images.

    Here's a youtube video of Moguri mod for FFIX (Ignore the terrible BGM, it's optional)

    Last edited by BURN_BABY; 4th March 2019 at 10:21 pm.

  6. Post
    #6
    BURN_BABY wrote:
    Very interesting to see the differences between the ML techniques as well. Check this comparison from FFIX out, for example. Megapixels looks closer to the original, even closer to ESRGAN trained with FFIX images.
    I'd totally like a comparison versus the source material before the compression to work on the PS1. Below are a few surviving assets/art, link to more below.

    Click image for larger version. 

Name:	11.FF9_IceCave.jpg 
Views:	27 
Size:	445.6 KB 
ID:	225361 Click image for larger version. 

Name:	rVltgrm.jpg 
Views:	25 
Size:	397.7 KB 
ID:	225362

    Click image for larger version. 

Name:	Du48mZR.jpg 
Views:	25 
Size:	134.2 KB 
ID:	225363 Click image for larger version. 

Name:	KZjqOw6.jpg 
Views:	23 
Size:	124.3 KB 
ID:	225364

    More at : https://www.neogaf.com/threads/the-l...thread.551612/

  7. Post
    #7
    XSpike wrote:
    I'd totally like a comparison versus the source material before the compression to work on the PS1. Below are a few surviving assets/art, link to more below.
    This is where it gets very cool - this is source



    this is upscale



    Some guy has already upscaled all of FFIX's backgrounds, you can check them out here.

    Edit: I also think the light windows might've been fixed in the Moguri Mod also.

  8. Post
    #8

  9. Post
    #9
    i've only looked at the ff8 video so far but it doesn't look much different to the original.

  10. Post
    #10
    Looks good, just seems the AI sucks at improving the image when the camera is moving fast or even just moving really. Awesome on static images, it'll get better.

  11. Post
    #11
    This one is really good, if you've seen the original


  12. Post
    #12
    Dem feels when you watch some of these.... **** it I'm firing up an emulator this weekend

  13. Post
    #13
    XSpike wrote:
    Looks good, just seems the AI sucks at improving the image when the camera is moving fast or even just moving really. Awesome on static images, it'll get better.
    It was designed for single images, not video, so the chances of it getting better are pretty slim under the current implementation.

  14. Post
    #14
    Yeah but its also an AI it'll learn to better do motion or be implemented on to better do motion on a different version & learn from that. Its early days.

  15. Post
    #15
    BURN_BABY wrote:
    It was designed for single images, not video, so the chances of it getting better are pretty slim under the current implementation.
    AS XSpike says the exact opposite is true. It's machine learning so it will get better. The more it does and the more input from those using it will lead to it being far better far faster than anything we've seen previously

  16. Post
    #16

  17. Post
    #17
    ChrisB wrote:
    AS XSpike says the exact opposite is true. It's machine learning so it will get better. The more it does and the more input from those using it will lead to it being far better far faster than anything we've seen previously
    Just because 'it's machine learning' doesn't automatically mean it'll get better for a job it isn't even designed to do. It also seems like upscaling tech like this is also designed to be used in conjunction with other traditional Photoshop techniques to remove unwanted artifacting, rather than replacing Photoshop as a 'set and forget' solution, which is what people seem to be using this for in FMVs.

    Another thing to take into consideration is Compression artifacting, which only gets more noticeable as upscaling techniques are applied to the video.
    Last edited by BURN_BABY; 7th March 2019 at 1:11 pm.

  18. Post
    #18
    BURN_BABY wrote:
    Just because 'it's machine learning' doesn't automatically mean it'll get better for a job it isn't even designed to do. It also seems like upscaling tech like this is also designed to be used in conjunction with other traditional Photoshop techniques to remove unwanted artifacting, rather than replacing Photoshop as a 'set and forget' solution, which is what people seem to be using this for in FMVs.

    Another thing to take into consideration is Compression artifacting, which only gets more noticeable as upscaling techniques are applied to the video.
    Your entire post is problematic due to incorrect assumptions. The progress generative adversarial networks has been making does include motion video along with massive leaps in quality because that's what it is supposed to do. Deepfakes are one result of the technology being used outside its original planned scope.
    The technology is being used for things right now it was not initially designed for and that's the point. the fact you're calling it "upscale tech" seems to indicate you don't understand what it is at all. I admit my knowledge here is extremely limited and I am only just getting my head around the machine learning aspects. But as to your point that "doesn't automatically mean it will get better" that's not really true as it is by design supposed to learn and "get better". How people use it will also improve over time. I am excited to see how it will be used in the future and how it could be used in other mediums such as film restoration, or even potentially reconstruction of damaged visual media. Who knows where this could go??

    On a related note

  19. Post
    #19

  20. Post
    #20
    ChrisB wrote:
    Your entire post is problematic due to incorrect assumptions. The progress generative adversarial networks has been making does include motion video along with massive leaps in quality because that's what it is supposed to do
    Err no, it's ESRGAN is a method of single image super resolution, and yes, it upscales from a low resolution image. The way people are doing FMVs is through 'brute forcing' the FMV into individual frames and running ESRGAN through each frame - however, applying super resolution to each frame in this way results in terrible compression artifacts, for example;



    At a guess, this is largely to do with how the video was originally encoded (using a lossy format) and the size restrictions of PS1 games, particularly those that have to fit on a single disc vs. 4 disc RPGs, for example. ESRGAN isn't designed to overcome these compression artifacts, meaning that somebody would likely have to go frame-by-frame and manually edit these artifacts out.

    It seems the optimal conditions for ESRGAN/Gigapixel et al. to work are a single, uncompressed low-resolution image. Put simply, if it worked well for videos, don't you think that would be documented on their paper?
    Last edited by BURN_BABY; 7th March 2019 at 7:29 pm.

  21. Post
    #21
    Still its early days, the program was designed for a stall image but if its programed to know that its apart of a video with frames before & after it'll match them together. It'll just learn that the camera is moving fast so it'll use a different technique for these frames, where on a stall image it just doesn't have to worry about it.

    Just easier to start with a stall image, before creating an AI designed to handle a video with frames before & after that should result in those images closely matching each other taking away these artifacts. At the moment its would be treating each image as its own thing resulting in these artifacts being placed in different locations each time making it really off putting for us.

    Still after many generations of it handling these videos & people selecting the better image each time, it'll improve until those artifacts are no more on each frame, putting the video together would result in a perfect vid as each frame matches each other.

    Either way the AI should with human input at the end of each gen telling it what is best, should solve these issues.

  22. Post
    #22
    BURN_BABY wrote:
    Err no, it's ESRGAN is a method of single image super resolution, and yes, it upscales from a low resolution image.?
    And I am not talking about a single technique being used so, yes. And once again these will improve because that is what they're designed to do.

  23. Post
    #23