QOI - The “Quite OK Image Format” for fast, lossless image compression

[wilberforce]

@phoddie

This looks like a great fit for Moddable.
https://phoboslab.org/log/2021/11/qoi-fast-lossless-image-compression

https://qoiformat.org/
https://github.com/phoboslab/qoi

QOI is fast. It losslessy compresses images to a similar size of PNG, while offering 20x-50x faster encoding and 3x-4x faster decoding.

QOI is simple. The reference en-/decoder fits in about 300 lines of C. The file format specification is a single page PDF.

The O(n) decoding would suit the micros and might speed up animations.

Someone has done (an out of date) typescript port:
https://github.com/panzi/jsqoi

[phoddie]

Interesting find. Some notes:

• It looks like it could offer a nice balance of feature for some situations.
• It should be possible to integrate with Poco rendering, though it would mostly be a rewrite of the decoder since the reference implementation isn't a streaming decoder.
• Clipping requires parsing through all pixels to the top and left of the clip origin, so not optimal for an atlas of UI elements.
• It uses 8-bit per channel to store colors. That will work, but most of our targets are 565, so each fully specified color uses one more byte of data than will ever be displayed. A variation of the image format could optimize that.
• It won't be faster than what's running on device now for UI rendering (we would never try to render a PNG directly -- both because of speed and memory - except in exceptional situations where PNG compatibility is the priority.

[phoddie]

Thinking about this some more... the compressed bit stream is really tuned for pixels that have 8-bits per component. For example, the big delta packet QOI_OP_LUMA is 16-bits, 14 of which are the payload. For 16-bit pixels, that is just 2 bits short of an entire pixel. The small delta packet QOI_OP_DIFF adjusts the low two bits of each color component. When displaying an 8-bit per component image on a 565 display, those bits will never be visible. And, of course, the color packet QOI_OP_RGB is 8-bits per component, which is one byte bigger than needed.

So, if used as-is, the main compression benefits are from constant color runs and using QOI_OP_INDEX to reduce recently used pixels to 8 bits. The benefit depends on the image, but it seems likely to deliver around 50% compression (better for images with solid color runs).

A 16-bit 565 variant (or 555) could do better:

• QOI_OP_RGB would save one byte
• QOI_OP_DIFFcould be re-biased to get the low bits of the components
• QOI_OP_LUMA might still be useful (once re-biased) but feels less so.

Do you know if someone has already investigated applying this to 16-bit pixels?

[wilberforce]

Do you know if someone has already investigated applying this to 16-bit pixels?

Sorry no - I just saw this as a news item.

There is also this Experimental branch: phoboslab/qoi#71

We could raise an issue as a starting point for discussing a 16 bit version.

[phoddie]

Asking about it in their repository makes more sense than just starting to hack at the problem. ;) Would you mind doing that?

[wilberforce]

@phoddie
New kid on the block - supports 565.

https://github.com/bitbank2/SLIC

[phoddie]

I modified the SLIC reference implementation to handle palettes in a way that is faster and more memory friendly for RGB565 rendering. It can also be smaller, as it only stores the actual number of colors used rather than always reserving space for 256 colors. It isn't intended a complete solution, but just enough to experiment with.

Using that I was able to encode images to use to implement 8-bit decoding. That works. I need to do tests to determine how well it performs, especially compared to RGB565 SLIC images.

[phoddie]

Did some more work on SLIC, this time to support alpha channel better. Detailed notes are available.

[wilberforce]

Have you sent the author your proposal? As his target is Arduino code he could be open to updating....

[phoddie]

Not yet. I'm planning to open an issue on that repository.

[phoddie]

Here's the issue. The notes are updated as well.