Anton Firszov

> What's the point of these tests? They look like local debugging tests to [report difference file-by-file](https://github.com/SixLabors/ImageSharp/blob/d81c511f068ac0fdb580310fe0d2328ca713e919/tests/ImageSharp.Tests/Formats/Jpg/SpectralToPixelConversionTests.cs#L61-L62), we can delete them.

> Tests from `DecodeBaselineJpeg` group do the same but they only log the difference if it's over the threshold. That's what we need normally, I probably added `Decoder_PixelBufferComparison` to have...

> leave rounding in place (I would prefer not to), make test pass difference threshold higher or re-encode reference images. I agree now that we should re-encode the images with...

> can you do diff command on ImageMagick with initial testorig.jpeg and generated png from ImageMagick? I don't know how to do diffs with ImageMagick, so there should be a...

Raising threshold can be an acceptable temporary solution, though I also consider it to be a tech debt. The long term thing would be to have a reference image generator...

> But I still don't understand why no-rounding approach produces worse results in the first couple of images Wait, aren't results "worse" only in compared to libjpeg and WIC runs...

I see, missed that removal of rounding also increases difference to PS reference. I won't exclude than that PS also uses integer algorithms, and would accumulate some generational loss the...

> Focusing on Jpeg decoding optimization for now I would advocate for being as radical as possible. > I propose redefining the entire JpegPostProcessor pipeline as three separate implementations based...

@michasacuer great news! Thanks! 👍

@SaltyDH cool that you are interested in this! :) Unfortunately, these classes are kinda messy, and need a lot of refactor before making them public. We can only take care...