How My Images Are Dithered
dead.garden
dead.garden
Our paper was about 3.5 meters wide, and a single roll length was 2–3 kilometers which ran for 20-30 minutes. The process delivered high quality at high speed, so we printed titles like Playboy and the IKEA catalogue in every language (for a language change we only had to swap the black cylinders, since the data was optimized so the text was mostly single color black). A print run of the IKEA catalogue, around 40 million copies, took about two months. It was brain numbing but also really zen. I miss it sometimes.
It aims to 'knock out' some areas of the image: make them transparent in the digital image so those areas are not part of the final film. This means less of the surface of the t shirt has film on it, allowing it to be breathable and flexible.
I got some DTF prints made before I knew you were meant to do this. The tutorials online required a lot of steps so I made the above for next time. Not tested for real yet.
Very important clarification, there.
Yeah, if you're using lossless compression like PNG or GIF, adding a lossy step to reduce the number of colors in the palette can help.
If you're compressing photos, you should use JPEG or WebP.
Anecdotally, JPEG usually beats dithering, and WebP always does.
It's fine to dither images but please please admit that you're doing it for the aesthetic. WebP has very broad browser support and absolutely blows dithering out of the water for compression. You are not going to beat years of research and tuning with "use fewer colors".
I remember a YouTube video about "How we did the impossible and got full-motion video for this Sonic game on a system that couldn't do FMVs", and the answer was like, "Well, we used a low resolution, and then we reduced the color palette, and then we cut the frame rate, and then we did interlacing." None of that's impossible! That's like saying you built the Pyramid at Giza using only a bucket of Lego bricks! It's not the same pyramid! Anyone can use a lower resolution! It takes effort and teamwork to come up with a codec that actually compresses data without artifacts!
They did in the article.
> My goal was to immitate a printed image.
A dithered image almost certainly beats a non-dithered JPEG compressed image in terms of file size, because you can JPEG compress the dithered image. Thus, if you like the look of a dithered image, you get file size reduction as a bonus.
I swear, some people on HN intentionally misinterpret and misrepresent the text of an article just to argue against it in bad faith with these smarmy "um actually" type arguments. For some reason I do not understand, Hacker News tends to surface these comments near the top of its comment sections.
That's backwards. Dithering (and simulated halftoning, as in this case) adds detail/noise, making images harder to compress efficiently.
For an algorithm like JPEG, if anything, you'd need to blur rather than dithering.
Depending on the image context it can also backfire, though not so much for photos, but rather screenshots. E.g. sometimes a 2 color dithered image results in a much bigger PNG than, say 16 colors without dithering, because bigger areas of uniform color compress much more efficiently. So I usually just try "everything", to find some sweet spot between file size and quality.
Is there a tool that allows you to select which area of the image to use what dithering parameters and color count for, with a live preview? I know I'll vibecode one myself just to see what it looks like, but maybe this already exists in a "real" application? Then you could e.g. use dithering on the faces, but for the background just a few colors without dithering.
Edit: found it. One picture taken from up close. Other pictures the whole image when laid out over multiple pages in the software https://imgur.com/a/OeaIWuY
I would not have clicked on “How My Images Are Dithered” but this truncated version made it sound more like an interesting bug story.
You can go down a rabbit hole of dithering algorithms too, like Floyd Steinberg[1]
[0] https://github.com/makew0rld/didder [1] https://en.wikipedia.org/wiki/Floyd%E2%80%93Steinberg_dither...
>But if you care about the method this is not dithering.
> I looked it up, the translation in this case is, in fact, simply "half tones" :^)
Surface-stable fractal dithering:
https://www.youtube.com/watch?v=EzjWBmhO_1E
And the technical deep dive:
To learn more about halftones, I'd recommend PDF specification, pages 303-307.
https://opensource.adobe.com/dc-acrobat-sdk-docs/standards/p...
> Consider that a grayscale gradient can only have 256 distinct levels.
This is true, but also keep in mind that for print and low dynamic range displays we humans only need approximately 256 levels, as long as the levels are perceptually spaced out evenly. 8 gamma corrected bits per channel actually made sense for a long time, for both print and for TVs & monitors, until HDR displays came along, relatively recently. Needing dithering for 8 bit images displayed on a 10 bit display is perhaps less surprising than needing dithering even when your display and your printer are both only 8 bits, right?
HDR is what changed - it requires more than 8 bits per pixel, by definition!
Again, if you only have LDR (low dynamic range), the human visual system’s “just noticeable difference” or JND in the perception literature demonstrates we only need 8 bits when looking at things under ‘indoor’ lighting conditions. You don’t need dithering if you use LDR on a single specific 8-bit display. You do need dithering when either moving to a different display than the one you started with, or using an HDR display - capable of displaying more than 8 bits of color resolution.
Yes. Photoshop does the correct thing when going 8bit via the "Mode" menu. On the other hand, if you do "Export As → PNG" and check the "Smaller file (8-bit)" checkbox, it does NOT dither, and it produces banding artifacts.
Like ~99% of images you encounter on the Web.
Your dithered "truedinner.webp" has 189.166 bytes. The original image has 34.682 bytes with JPEG.
https://show.quicky.club/results/b5/30/12/16f36013492e06b658...
I tried doing something similar a couple of years ago and gave up after running into the exact same issues in the post. It's really difficult to replicate proper halftoning on a computer screen without getting moiré patterns and weird artifacts because of PPI, aliasing, and other stuff. Beyond that, print and screens are just fundamentally different ways of rendering things.
It has some fun animation options and settings to play with, and there's other dithering algorithms to play with.
Unhalftoning is called descreening.
I use this:
> This image features a close-up abstract scene styled with a classic halftone (Ben-Day dot) pop-art filter ...
https://share.gemini.google/O5eqZOKxwzw3
> This image is a halftone/thermal print styled photograph. The photo features a group of four people posing closely together indoors or in front of a windowed backdrop:
Why do you believe training on dithered images would learn how to see through any other kinds of image modification? In my experience, image NNs learn only what you explicitly show them, and they don’t learn concepts they’re not trained on. All the denoisers I’ve ever seen have trouble handling subtle changes in pixel filtering they haven’t seen in the training data, and they don’t easily learn how to denoise things like clouds or reflections. I’d be willing to bet that if you trained on halftoned images, NNs would not be able to handle Floyd Steinberg dither. They just don’t generalize automatically.
NNs can certainly learn beyond what they were trained for as new architectures have demonstrated. See benchmarks for any if the dan hendryks imagenet corruption dataset benchmarks (imagenet c r and a). Data + architecture work together to enable.
I have work using neural networks trained on synthetic aperture sonar SAS imagery using language only. The SAS datsets are extremely hard to come by. The dataset I used was from parts if world I can assure you we're never in the training set. My approach exceeds what clip models can do zero shot wise. Therefore, the network classified imagery if a modality and environment 100 percent not guaranteed to be in the training data.