JPEGXL lossless float16 is not lossless

[Skielex]

Hey,

First of all, really nice work on supporting so many image formats. I've been playing around with lossless encoding of some float16 data. However, when decoding there appears to be artifacts. I don't know if the issue is in your wrapper or in libjxl, however, before I raise an issue over there I just wanted bring it up here, since it's probably easy for you to check if the issue resides here.

The issue can be replicated on WSL2 Ubuntu Python 3.10 with imagecodecs==2024.9.22 using this test:

np.random.seed(42)
r_f16 = np.random.random((128, 128)).astype(np.float16)
r_f32 = r_f16.astype(np.float32)

r_jxl_f16 = jpegxl_encode(r_f16, lossless=True)
r_jxl_f32 = jpegxl_encode(r_f32, lossless=True)

r_jxl_f16_dec = jpegxl_decode(r_jxl_f16)
r_jxl_f32_dec = jpegxl_decode(r_jxl_f32)

np.testing.assert_equal(r_f32, r_jxl_f32_dec)  # Passes
np.testing.assert_equal(r_f16, r_jxl_f16_dec)  # Fails

Visually, it looks like this:

[cgohlke]

I can reproduce this on Windows, AMD64 and ARM64. It looks like a bug in libjxl to me. The glitches are quite rare though, maybe one in several thousand random values.

Try to reproduce this with the cjxl/djxl tools and open an issue at https://github.com/libjxl/libjxl/issues.

[Skielex]

Ok, glad to hear it's not just me. I'll try the tools.

[Skielex]

The cjxl/djxl tools don't appear to currently allow any formats that support float16, so I haven't been able to find a way to reproduce the issue using the tools.

The issue appears to be related to specific input values. For instance, the input value 3.07e-05 is consistently decoded as 3.6e-07.

[Skielex]

I tested all 65536 possible float16 values. It appears that 3070 (~4.7%) of these are changed by encode-decode. Interpreted as uint16, the four affected regions are: [(512, 1024), (31745, 32768), (33280, 33792), (64513, 65535)].

If we plot which bits are turned on, it looks like this:

[cgohlke]

currently allow any formats that support float16

Did you check EXR? Also,the GIMP plugin supports float16. Not sure, I have not tried either.

[Skielex]

Looks like EXR support has been dropped for the time being:

cjxl.exe INPUT OUTPUT [OPTIONS...]
 INPUT
    the input can be JXL, PPM, PNM, PFM, PAM, PGX, PNG, APNG, GIF, JPEG

libjxl/libjxl#1662

Gonna try GIMP.

[Skielex]

Tried GIMP 2.99.19, which supports JPEG-XL. However, support appears a bit buggy.

Opening float16 or float32 jxl-files encoded using imagecodecs appears to work. float16 files contain the same artifacts that appear when decoding with imagecodecs, and float32 files appear correct, as when decoding with imagecodecs.

GIMP does not appear to support saving JPEG-XL as floating types.

I've already spent too much time on this, and there doesn't appear to be any good way of encoding float16 using libjxl outside creating my own set of bindings. I'll report the issue over at their repo and see what they think.

[kmilos]

FWIW, darktable can do fp16 EXR and TIFF. It should also decode any JXL.

For JXL encoding, 4.8.1 does just lossy floats, only the dev version can do fp32 lossless but crashes for fp16 lossless (also reported to libjxl; seems like there is no built-in fp32->fp16 down-conversion like there is when targeting integer lossless...).

[kmilos]

Looks like EXR support has been dropped for the time being

I have it here (MSYS2 package). Just means it was built w/o OpenEXR present for you. Doesn't mean it works though, haven't tested...

$ cjxl.exe
JPEG XL encoder v0.11.0 0.11.0-1 [AVX2,SSE4,SSE2]
Usage: C:\msys64\ucrt64\bin\cjxl.exe INPUT OUTPUT [OPTIONS...]
 INPUT
    the input can be JXL, PPM, PNM, PFM, PAM, PGX, PNG, APNG, GIF, JPEG, EXR

[Skielex]

@kmilos , thanks for chipping in. At this point, I'd be surprised if the error isn't in the encoder. Especially given the encoded size of a float16 256x256 image containing all 65536 possible values being only 826 bytes, while the encoded float32 image with the same values is 92 kb. I've attached the files in a zip.

images.zip

If you want, you're very welcome to try and open the f16.jxl file in Darktables and see if it's correct, or has the some artifacts. Then again, if you rely on libjxl, we wouldn't really learn anything new.

The zip also contains two EXR files that should contain the same 65536 values as the JXL files. If you've got cjxl/djxl working to EXR, perhaps you can check if they encode correctly.

[kmilos]

The zip also contains two EXR files

Those two are actually (uncompressed) TIFFs, not EXRs, and cjxl unfortunately doesn't support TIFF input...

In any case, renaming those to f{16,32}.tif and importing into darktable does show only f16.jxl indeed differs from the other three; note also the difference in the histograms attached.

f16.jxl (compensated by -17.1 EV, edge is further down, unexpected hole in histogram):

f16.tif (compensated by -15.5 EV, continuous histogram as expected):

P.S. The bottom halves of the image are black as dt discards/clips the negative values (which begs the question why the border is not exactly at the half height for f16.tif and f32.{jxl,tif}? I guess inf/NaN signaling take up that much space - 1024 values or 4 rows in your 256x256 test image?)

[kmilos]

Looking at the subnormal values w/ +11 EV compensation there are obvious differences as well, so there's definitely something going on w/ just the subnormal and inf/NaN representations...

f16.jxl w/ +11 EV:

f16.tif w/ +11 EV: