farbfeld

farbfeld is a lossless image format which is easy to parse, pipe and
compress.
It has the following format:

| Bytes | Description |
|--------|------------------------------------------------------------|
| 8 | "farbfeld" magic value |
| 4 | 32-Bit BE unsigned integer (width) |
| 4 | 32-Bit BE unsigned integer (height) |
| [2222] | 4⋅16-Bit BE unsigned integers [RGBA] / pixel, row-aligned |

http://tools.suckless.org/farbfeld/

>>20

Okay, test results are in:

8.1M Jan 13 12:53 hurl.ff
4.1M Jan 13 12:54 hurl.ff.bz2

8.1M Jan 13 12:56 hurl_noalpha.ff
3.6M Jan 13 12:56 hurl_noalpha.ff.bz2

Where hurl.ff is basically IMAGE1, a random set of RGBA values.
hurl_noalpha.ff is IMAGE2, the version with FF alpha values.

What I did is the following (for reproducibility):
I created a new image in GIMP, 1024x1024 and applied the Hurl-filter (which basically randomizes all RGBA values) (yes it's really called like that :P).
Then I wrote a little program to only set the Alpha value to 65535 (32 bit unsigned max value), yielding in hurl_noalpha.ff.

Now let's compare the values:
The uncompressed size obviously didn't change. But let's compare the bzip2-versions, which surprisingly still found some way to compress the data to half.

3.6/4.1 ~ 0.88 -> bzip2 managed to recognize the Alpha channel and saved 12% of space in that process.

Given this is a statistical issue, the tests could be repeated multiple times (with a lot of random data). However, we see the trend here. Bzip2 is not an ideal compression algorithm, however, 12% of maximum 25% is actually quite impressive.

And I don't know what this alleged AI compression research expert wants. It's quite obvious that being able to predict upcoming data patterns may be the key to even better compression.
On the other hand, if he really is a compression researcher, it isn't surprising that this field is not dramatically proceeding. :P

>>21

It's the network byte order. If this goes beyond your small head it's not my problem.

>>19

troll troll troll :P

>>25

Okay, here are the results

6.1M Jan 13 13:25 hurl_48bit.ff
3.1M Jan 13 13:25 hurl_48bit.ff.bz2

I cut out the Alpha value, so it only stored RGB-pixels. Now, as we can see, this gets close to 25% based on the original, which means that we properly assessed the test's circumstances.