Strange color issue happen on -q 5~6(--quality)

[Lee-lithium]

Hello Ilya Kurdyukov, Sorry to bother you.

I get some strange color issue on non-photo drawing content,
those issue will happen on -q 5~6(--quality).

In this non-photo picture case, green book middle area have some tiny red pixel,
-q 3~4 haven't happen this issue.

original-cut-png (jpeg q90 420)

qs-q6-png

JPEG Quant Smooth : Copyright (C) 2020 Ilya Kurdyukov : 2020-05-17
Build date: Jan 11 2021
https://github.com/ilyakurdyukov/jpeg-quantsmooth/releases/tag/1.20210110

jpegqs64 -q 6 -o -t 12 test.jpg test-qs.jpg
sample.zip

[ilyakurdyukov]

I'm not really in the mood to solve this now because of the current world situation that happened because of my country.

This is when the UV upscaler makes the wrong decisions. Need a way to limit this. As I see it, mostly the upscaler works correctly, but in some places it creates artifacts.

Try replacing the text "16.0f" (use "search and replace") in quantsmooth.h with a smaller value (8.0f, 4.0f, 2.0f). Maybe I chose too high value for the limit.

[Lee-lithium]

I understand, Thank you for your reply.
I'm sorry about this world situation, hope everything’s gonna be okay.

[ilyakurdyukov]

If you find a better value for this constant - feel free to tell me.

[ilyakurdyukov]

This limit doesn't solve the problem, need a different solution.

[zvezdochiot]

Hi @ilyakurdyukov .

Maybe track the value of the coefficients on a given image via gdb?

Or use #if DEBUG for dump?

[ilyakurdyukov]

It seems to be the sum of small errors from different methods. In the original, the chroma has a red dot here, which gets brighter after reducing the artifacts, and then colors part of the line after the chroma upscale and gets even brighter. Unfortunate artifact from recovering the lost precision.

I saw this once in my test data as a miscolored dot.