The 5 Errors of Dandere2x 3.7.2 (3.3.1 is better?) about dandere2x HOT 39 OPEN

You don't realize how valuable this feedback is for me. I'll use this as a reference point on how to fix current dandere2x as well as drawing inspiration from the new dandere2x.

Seriously, thank you. I'm very impressed by the effort you put into this.

from dandere2x.

Thank you for your input. I'll research this tonight

from dandere2x.

What settings were these used under?

from dandere2x.

@akai-katto I used maximum block and maximum image quality.

from dandere2x.

Hm, bigger block size typically is worse, I wish I had specified this before you went into it. Thanks though.

Curious to see if smaller block size (say 20?) has any improvement for your results?

from dandere2x.

@akai-katto I don't like to use low blocks, because you can easily see the grid. For example, lines, using a low block makes them more noticeable. But I will try it now.

from dandere2x.

@akai-katto I have tested 3.7.2 with block 20 and it is still the same.
I uploaded the video of the test in Google Drive in the folder "Other Tests".

from dandere2x.

I got the same results.

I'll take a break from dandere2x rework to work on this.

from dandere2x.

https://drive.google.com/file/d/1Vq7QvMOdn7yk1PWLYX5QPniD3OEPlx8v/view?usp=sharing

Would you say this looks less noisy, or the same (this is dandere2x rework WIP)

from dandere2x.

@akai-katto A great improvement to the contrast which remains the same as 1.3.1! although the color still varies from the original tone, which I think must be difficult to correct.

The noise remains exactly the same. In the final part of the video is the key to see the noise. Or as I called it (spots). Maybe it would be better to use the 60 block, it's the same as the 20 but easier to identify.

from dandere2x.

how do you edit the photo to dramatize the spots? I can barely notice them on my VA monitor, so I'm trying to work around replicating it locally

from dandere2x.

here's blocksize 60 on new dandere2x, if you want to share your thoughts

https://drive.google.com/file/d/1dq7fQsWbNsmA16LyTSuv27Ts3MBjq_7V/view?usp=sharing

from dandere2x.

here's blocksize 60 on new dandere2x, if you want to share your thoughts

https://drive.google.com/file/d/1dq7fQsWbNsmA16LyTSuv27Ts3MBjq_7V/view?usp=sharing

Access denied xd

from dandere2x.

shared

from dandere2x.

@akai-katto I think the best program to see the noise is "avidemux", using a filter: Artistic > Dynamic Threshold.
I don't know why I had not used this program before for testing. [Block 60]

from dandere2x.

Does the input video demonstrate the same noise? I'm wondering if it's an input video error or dandere2x error. Hoping it's an input video error but we'll see.

from dandere2x.

@akai-katto This is what the original video looks like. It is noticeable that the noise is different, since in this one it is quite small, while the Dandere2x one is mostly grid-based.

from dandere2x.

Would you mind trying the new dandere2x prototype out, and leaving your thoughts on it? https://www.reddit.com/r/Dandere2x/comments/zctvfo/dandere2x_rework_proof_of_concept_ready_for/

from dandere2x.

@akai-katto I'll try it now, I only use Microsoft Defender so hopefully it won't be a problem xd.

from dandere2x.

@akai-katto At the moment it strikes me that it is using the integrated graphics xd, maybe I should start the program with the Nvidia manually later.

from dandere2x.

let me fix this

from dandere2x.

https://github.com/akai-katto/dandere2x-python-rework/releases/tag/0.0.2

You should be able to download the latest versioning allowing selecting GPU in 0.0.2.1.

Simply goto your config_files, open up output_options.yaml, and change the following variables

waifu2x_ncnn_vulkan:
  model_name: "models-cunet"
  tile_size: 200  # note: if you get black images try lowering this tile argument to some lower value.
  pre_padding: 18
  client_ports:
    client0:
      gpuid: 0
      receive_port: 43560
      send_port: 43561
    client1:
      gpuid: 0
      receive_port: 43562
      send_port: 43563
    client2:
      gpuid: 0
      receive_port: 43564
      send_port: 43565
    client3:
      gpuid: 0
      receive_port: 43566
      send_port: 43567

->

waifu2x_ncnn_vulkan:
  model_name: "models-cunet"
  tile_size: 200  # note: if you get black images try lowering this tile argument to some lower value.
  pre_padding: 18
  client_ports:
    client0:
      gpuid: 1
      receive_port: 43560
      send_port: 43561
    client1:
      gpuid: 1
      receive_port: 43562
      send_port: 43563
    client2:
      gpuid: 1
      receive_port: 43564
      send_port: 43565
    client3:
      gpuid: 1
      receive_port: 43566
      send_port: 43567

from dandere2x.

@akai-katto Ok I'll try it, but I'll wait for this rendering to finish, it's 15 minutes left.

from dandere2x.

@akai-katto The result was the same as 3.7.2 with the same amount of errors, but slower. Now I will test the version with the GTX1050 GPU.

from dandere2x.

@akai-katto The result was the same as 3.7.2 with the same amount of errors, but slower. Now I will test the version with the GTX1050 GPU.

what quality setting did you use?

from dandere2x.

@akai-katto Everything to the maximum as always xd, because this way you can see the errors directly. Because whenever I use small blocks they are the same errors but in smaller boxes. (scale x2)

PD:now I will do it with block 20

from dandere2x.

@akai-katto denoise 3, scale 2x, block 20, image quality 100, 2 instances, time 6:21, same issues.
The only major noticeable difference is the large reduction in CPU usage.

from dandere2x.

@akai-katto Everything to the maximum as always xd, because this way you can see the errors directly. Because whenever I use small blocks they are the same errors but in smaller boxes. (scale x2)

PD:now I will do it with block 20

https://github.com/akai-katto/dandere2x-python-rework/releases/download/0.0.2/uncapped_quality.exe

i uploaded "uncapped_quality.exe" which allows you to set quality up to 50k (lol).

I'm curious if you'd find the quality issues are fixed for you if the quality is set high enough.

from dandere2x.

@akai-katto denoise 3, scale 2x, block 20, image quality 100, 2 instances, time 6:21, same issues. The only major noticeable difference is the large reduction in CPU usage.

is it slower than 3.7.2 or 3.0.3?

from dandere2x.

@akai-katto I'm going to test with 4 instances, in addition I count 20 extra frames here than the tests.

from dandere2x.

@akai-katto denoise 3, scale x2, block 60, image quality 100, 4 instances, time 06:48, same issues.
And unfortunately, another problem has been added. The video feels like it sticks in some parts.
It's probably something about using 4 instances, since I had never used them, probably this bug is also in previous versions.
By using only 2, the problem was reduced by half, surely this does not happen with only one instance. I will do this test now with only 1 instance and so I will equalize the conditions of all the tests.

from dandere2x.

darn, i thought i fixed the "sticking" parts. I'll upload another binary, can you let me know if it fixes it? I won't ask any more from you tonight, as it's getting late.

from dandere2x.

https://github.com/akai-katto/dandere2x-python-rework/releases/download/0.0.2/0.0.2.2.exe

from dandere2x.

@akai-katto I have no problem with the time, first I will test the version you had already sent with a single instance and see how it performs, then I will look at the new one.

from dandere2x.

On my end, the binary I sent you fixes the sticky issue.

from dandere2x.

@akai-katto In the previous version with a single instance it is solved checked.
denoise 3, scale x2, block 60, image quality 100, 1 instance, time 08:20, same issues.
Although this new Dandere2x I do not know if it has corrected the dots because I have not seen any, but I would have to do more tests.
Now I'm going to test the new version with 2 instances

from dandere2x.

Thanks for testing btw, I'm relying on the community to catch bugs like this.

from dandere2x.

@akai-katto denoise 3, scale x2, block 60, image quality 100, 2 instances, time 06:51, same issues 3.7.2
Sticky video corrected.
It's a big improvement with version 3.7.2 in optimization issues, it has lowered my processor temperature by 10°C.

The time considered in the other tests goes from frame 10 to 350, this one too, but considering that, in the previous tests they didn't take the time of the frame extraction or whatever happened before that I don't understand much xd.

The most serious error that I see now are the "spots" or "noise" that is very noticeable.

@akai-katto compared to 3.7.2 which had a lot of dots, I think it was corrected in this version. So it would be the version with less bugs, although as the "spots" or "noise" are so exaggerated, I will give it a 4.

I left the test there if you want to see the differences.

The color remains the same as in 3.7.2.
I would like it to be corrected like the 3.7.2 correction of 2 days ago.

https://drive.google.com/file/d/1Vq7QvMOdn7yk1PWLYX5QPniD3OEPlx8v/view?usp=sharing

Would you say this looks less noisy, or the same (this is dandere2x rework WIP)

from dandere2x.

The progress of the new Dandere2x is incredible.
"Quality Coefficent": If this is low it generates the problem of the dots. So, by having it at 100, this is solved.
Lines still appear depending on the configuration
The processor heats less, the error of the dots, lines, color was corrected, the only thing that remains is the noise.
With this configuration I have managed to reduce everything as much as possible.
Test_[Cun][w2][s2][d3][q100][b60][t050]

It is already too late, tomorrow I will take a closer look at the tests and see which one is better in each area.
https://drive.google.com/drive/folders/1ZxyjDErgJGih6EyGzMY_QBjTN2gwtMLV?usp=share_link

from dandere2x.