Fighting JPEG color banding

Fighting JPEG color banding(uploadcare.com)

136 points by igoradamenko 3 years ago | 52 comments

I remember that almost 20 years ago, I played with a utility that allowed manual JPEG optimization by painting regions with desired quality settings. I think it was included on a CD bundled with "Magazyn internetowy WWW" [1]. Does anyone remember such program?

[1] https://sprzedajemy.pl/www-magazyn-internetowy-2003-nr-04-03... - incidentally this must be the very issue, see the teaser about image optimization

lynguist 3 years ago |

This is OK for icon sized images, but it hurts me when I read webpages that incorporate photographs or visualizations in low resolution. They keep saying “Web resolution” and talk about something like 200x200 images maximum.

Even for the past 20 years, when I read webpages with images, I inspect them closely. Often it’s something like a newspaper article with photographs or Wikipedia. I specifically set my Wikipedia settings to deliver me images at the highest possible resolution and now I can actually enjoy reading it. Specifically I use the Timeless skin and set the thumbnail size to the maximum.

Sometimes I come across webpages that describe something like historic trains and all they have are icon sized photographs of them. It’s so sad.

viraptor 3 years ago |

Thing I'd love to try if I had time - a compressor which derives the best table for the image. I'm imagining a loop of: start with the default, compress, calculate the difference from the original, dct the error to see which patterns are missed, adjust the table, repeat. Stop on some given error/size-increase ratio. (yes, I'm trying to get someone else nerd sniped into doing this)

Edit: something like this https://www.imaging.org/site/PDFS/Papers/2003/PICS-0-287/849...

So from the older known ones there DCTune and DCTex methods, but it seems neither is available for download anywhere.

jncraton 3 years ago | |

Guetzli was already mentioned and roughly does what you are talking about.

MozJPEG [1] includes several quantization tables that are optimized for different contexts (see the quant-table flag and source code for specific tables[2]), and the default quantization table has been optimized to outperform the recommended quantization tables in the original JPEG spec (Annex K).

It's also worth noting that MozJPEG uses Trellis quantization [3] to help improve quality without a per-image quantization table search. Basically rather than determining an optimal quantization table for the image, it minimizes rate distortion on a per-block level by tuning the quantized coefficients.

Both the SSIM and PSNR tuned quantization tables (2 and 4) provided by MozJPEG use a lower value in the first position of the quantization just like this article suggests (9 and 12 vs the libjpeg default of 16).

[1] https://github.com/mozilla/mozjpeg

[2] https://github.com/mozilla/mozjpeg/blob/5c6a0f0971edf1ed3cf3...

[3] https://en.wikipedia.org/wiki/Trellis_quantization

homm 3 years ago | | |

> MozJPEG use a lower value in the first position of the quantization just like this article suggests

It has lower value in the first position of the base table, i.e. the table which is used for q=50. With lower qualities this value scales up. This delays color banding from q=50 to roughly say q=40, after that the same effect is appears.

viraptor 3 years ago | |

Turns out there is a project which does that: https://github.com/google/guetzli

duffyjp 3 years ago | | |

Guetzli is really hamstrung by its resource usage. When it first hit the news I tried it out, and compressing a full quality JPEG from my phone could take 20-30 minutes on an i7.

hansword 3 years ago |

I love this and would have dearly needed it like 5 years ago. Now, it is still a very interesting read.

But given what we have already seen from Nvidia on video compression [0], I think within the next few years, we will move everything to machine-learning-'compressed' images (aka transmitting a super-low-res seed image and some additional ASCII and having an ML model reconstruct and upscale it at the client side).

[0] https://www.dpreview.com/news/5756257699/nvidia-research-dev...

wongarsu 3 years ago | |

Most images are still JPEG (3 decades old) or PNG (2.5 decades old). Countless better formats have been developed, but with the exception of WEBP we are still using the same image formats that existed during the dot-com bubble. Ubiquity trumps improvements in image size.

Better encoders for JPEG or PNG are the main avenues how you can achieve improvements without compatibility problems, and I think that will stay true for another decade, if not more.

BiteCode_dev 3 years ago | |

I fear that we will get used to some image patterns feeling right instead of reality if AI gets involved too much.

dylan604 3 years ago | |

honestly, this scares this shit out of me.

lossy compression is one thing, but to just say that an ML model suggests making pixels like this vs a mathematical formula is totally different things.

dehrmann 3 years ago | | |

What if it's an ML model suggesting parameters for jpeg? It could still hallucinate to some degree, but it's also more limited.

lifthrasiir 3 years ago |

The original JPEG, retroactively named JPEG1, totally lacked loop filters and the quantization factor of DC coefficient matters much more than modern formats. As an example, libjpeg q89 is noticably worse than q90 because the DC quantization factor changes from 3 to 4 (smaller factor means less quantization thus higher quality), quite a big jump.

cabirum 3 years ago |

There's also webp and heif and png and svg, and I believe all the existing formats already solve the image compression problem. The difference of 18kb vs 22kb from hours of microoptimisations is frankly irrelevant given the rate of networks getting faster.

yboris 3 years ago |

Just use JPEG XL (aka .jxl) - https://jpegxl.info/

You can re-save a JXL image a thousand times without deterioration :)

rasz 3 years ago |

Does this mean default quantization tables were badly picked after all? And someone only noticed after 30 years?

It sounds like the employed solution only modifies first element from [16,17] to [10,16].

homm 3 years ago | |

It was not badly picked for regular use cases. Just, before the retina displays appeared no one was interested in extreme low bitrate and no one knew that different artifacts had different impact with high density.

At least this problem was highlighted by Kornel, one of the mozjpeg author here: https://github.com/mozilla/mozjpeg/issues/76

> the employed solution only modifies first element from [16,17] to [10,16].

Correction: 16 and 17 are values from the base tables, which means this table is used with q=50. With q=25 it will be [32, 22, 24, 28, 24, 20, 32, 28…] (in zigzag order). The employed solution is to always limit the first value by 10 regardless of q: [10, 22, 24, 28, 24, 20, 32, 28…]

mozjpeg chosen the different approach: it still scales all values based on q, but has significantly changed the default base table. It helps, but doesn't eliminate color banding completely (you can still see it on the example from issues/76).

GuB-42 3 years ago | |

The tables were good for the time they were picked. And the beginning of the article, it shows that at low resolutions, both ringing and banding are unacceptable. At high resolutions, beyond what was considered normal when the original quantization tables were chosen, then ringing becomes much less of a problem, so it makes sense now to change the quantization tables to prioritize banding over ringing.

It is no secret that high resolution images and low resolution images compress differently, and modern codecs are optimized for high resolutions in a way that older codecs weren't. For example going from H.261 to H.266 globally improve video compression at every step, but it is most apparent at higher resolutions.

hansword 3 years ago | | |

> At high resolutions, beyond what was considered normal when the original quantization tables were chosen, then ringing becomes much less of a problem

Slightly more precise: ringing becomes much less of a problem in high resolution images when viewed on a high resolution screen (in Apple language: retina).

erlndgdt 3 years ago |

How is this different from say, what https://kraken.io is doing. When I upload images there I get a smaller file size and the image looks no different

hansword 3 years ago | |

The difference is that the kraken 'about technology' website [0] gives no useful information as to how they are doing it (I guess it's 'proprietary information'), while this article gives a very detailed description of how to compress jpegs.

In other words: end users use kraken, developers read this article.

[0] https://kraken.io/about/technology

ntoskrnl 3 years ago | | |

I can take a guess at that "proprietary information". Most image optimization sites are just thin wrappers over something like mozjpeg/zopflipng/etc. 1% tech, 99% marketing

homm 3 years ago | |

It's for free.

saltminer 3 years ago |

> Ok, but where does this table come from when we need to save a file? It would be a big complication if you had to construct, and transmit 64 independent numbers as a parameter. Instead, most encoders provide a simple interface to set all 64 values simultaneously. This is the well known “quality,” which value could be from 0 to 100. So, we just provide the encoder desired quality and it scales some “base” quantization table. The higher quality, the lower values in quantization table.

I never really thought about how that "quality" slider worked (besides making the compression lossier), but it makes perfect sense now! It always amazes me how much I take for granted.

I always treat compression like a black box: "-crf 23" for H264, PNG and FLAC are nice but MP3 320s and 90+ "quality" JPEGs are good compromises, etc. And that's just for the stuff I deal with, there's no telling how much lossy compression goes on behind the scenes on my own computers, let alone all the stuff served up over the internet. There's so much lossy compression in the world, from EP speed on VHS tapes to websites reencoding uploaded images to every online video ever, it's crazy to think about.

Firadeoclus 3 years ago |

Though its impact may be limited, this is some nice work!

Now could someone look at how video codecs can produce excellent high-detail scenes and motion in 4k resolution while at the same time making a blocky mess out of soft gradients, especially in dark scenes with slow movement?

LordDragonfang 3 years ago |

If you want a similar read on DCT-based encoding, I highly recommend this article which lays it out in an extremely digestible format:

https://sidbala.com/h-264-is-magic/

mcdonje 3 years ago |

I ran into this problem exporting images from Dark Table. The solution I ended up going with was simply exporting as png instead of jpg.

hulitu 3 years ago |

just use PNG. JPEG sucks.

$ gm identify test.tiff test.tiff TIFF 6240x4160+0+0 DirectClass 8-bit 1.2Mi 0.000u 0m:0.000001s $ time gm convert test.tiff test.jpg real 0m0.282s user 0m0.193s sys 0m0.089s $ time heif-enc test.jpg -o test.heif real 0m1.901s user 0m22.960s sys 0m0.180s