10bit JpegXL is saved as 16bit file

[jmpfar]

Hi, really great library ☄️ I am trying to save a 10bit JpegXL using the library and for some reason I am getting a 16bit JpegXL file back:

EFFORT = 2

array = np.fromfile(file, dtype=np.uint16)
array = array.reshape(RESOLUTION)

outfile = Path("/tmp/blah.jxl")
data = jpegxl_encode(array, distance=0, effort=EFFORT, lossless=True, bitspersample=10, photometric="GRAY")

outfile.write_bytes(data)

The array itself is uint16 but stores integers in the range of 0-1023

$ jxlinfo /tmp/blah.jxl
JPEG XL file format container (ISO/IEC 18181-2)
JPEG XL image, 700x700, (possibly) lossless, 16-bit Grayscale
Color space: Grayscale, D65, Linear transfer function, rendering intent: Relative

If I'm reading this correctly, it is getting overriden here: https://github.com/cgohlke/imagecodecs/blob/91389e7e325fadb93b4c7e21e8284766d0ae10b4/imagecodecs/_jpegxl.pyx#L327-L329

Thanks!

[cgohlke]

The bitspersample parameter currently specifies the bit depth of the pixel input (set with JxlEncoderSetFrameBitDepth). The bit depth of the image file is still 16-bit (basic_info.bits_per_sample = 16). It should be possible to improve that.

[cgohlke]

I think you are right that basic_info.bits_per_sample needs to be set to bitspersample. What's missing was to also set bit_depth.dtype = JXL_BIT_DEPTH_FROM_CODESTREAM in the decoder such that the roundtrip test pass. Otherwise the decoder would scale from 12 to 16 bit.

The fix will be included in the next version.

[jmpfar]

Thanks! that could really help as right now my images turn out a bit too dark Another thing that I saw in the code that I couldn't fully figure out: https://github.com/cgohlke/imagecodecs/blob/91389e7e325fadb93b4c7e21e8284766d0ae10b4/imagecodecs/_jpegxl.pyx#L213-L220

it looks like bits_per_sample and exponent_bits_per_sample are being set but not used, even in the context of bit_depth

[cgohlke]

This seems to work:

diff --git a/imagecodecs/_jpegxl.pyx b/imagecodecs/_jpegxl.pyx
index 8592d59..1f64720 100644
--- a/imagecodecs/_jpegxl.pyx
+++ b/imagecodecs/_jpegxl.pyx
@@ -186,6 +186,7 @@ def jpegxl_encode(
         float option_distance = _default_value(distance, 1.0, 0.0, 25.0)
         size_t num_threads = <size_t> _default_threads(numthreads)
         size_t channel_index
+        uint32_t bits_per_sample
         bint is_planar = bool(planar)

     if data is out:
@@ -210,14 +211,13 @@ def jpegxl_encode(
     if not (src.dtype.kind in 'uf' and src.ndim in {2, 3, 4}):
         raise ValueError('invalid data shape or dtype')

+    memset(<void*> &bit_depth, 0, sizeof(JxlBitDepth))
     if bitspersample is None or src.dtype.kind == 'f':
         bit_depth.dtype = JXL_BIT_DEPTH_FROM_PIXEL_FORMAT
         bits_per_sample = 0
-        exponent_bits_per_sample = 0
     else:
         bit_depth.dtype = JXL_BIT_DEPTH_FROM_CODESTREAM
         bits_per_sample = bitspersample
-        exponent_bits_per_sample = 0

     out, dstsize, outgiven, outtype = _parse_output(out)

@@ -323,10 +323,14 @@ def jpegxl_encode(

     if dtype == numpy.uint8:
         pixel_format.data_type = JXL_TYPE_UINT8
-        basic_info.bits_per_sample = 8
+        if bits_per_sample < 1 or bits_per_sample > 8:
+            bits_per_sample = 8
+        basic_info.bits_per_sample = bits_per_sample
     elif dtype == numpy.uint16:
         pixel_format.data_type = JXL_TYPE_UINT16
-        basic_info.bits_per_sample = 16
+        if bits_per_sample < 1 or bits_per_sample > 16:
+            bits_per_sample = 16
+        basic_info.bits_per_sample = bits_per_sample
     elif dtype == numpy.float32:
         pixel_format.data_type = JXL_TYPE_FLOAT
         basic_info.bits_per_sample = 32
@@ -364,7 +368,7 @@ def jpegxl_encode(
                 basic_info.animation.num_loops = 0
                 basic_info.animation.have_timecodes = JXL_FALSE

-            framesize = ysize * xsize * basic_info.bits_per_sample // 8
+            framesize = ysize * xsize * dtype.itemsize
             if not is_planar:
                 framesize *= samples

@@ -595,6 +599,7 @@ def jpegxl_decode(
         JxlSignature signature
         JxlBasicInfo basic_info
         JxlPixelFormat pixel_format
+        JxlBitDepth bit_depth
         size_t num_threads = _default_threads(numthreads)
         size_t channel_index
         bint keep_orientation = bool(keeporientation)
@@ -619,6 +624,7 @@ def jpegxl_decode(

             memset(<void*> &basic_info, 0, sizeof(JxlBasicInfo))
             memset(<void*> &pixel_format, 0, sizeof(JxlPixelFormat))
+            memset(<void*> &bit_depth, 0, sizeof(JxlBitDepth))

             decoder = JxlDecoderCreate(NULL)
             if decoder == NULL:
@@ -760,9 +766,11 @@ def jpegxl_decode(
                                     ' not supported'
                                 )
                         elif basic_info.bits_per_sample <= 8:
+                            bit_depth.dtype = JXL_BIT_DEPTH_FROM_CODESTREAM
                             pixel_format.data_type = JXL_TYPE_UINT8
                             dtype = numpy.uint8
                         elif basic_info.bits_per_sample <= 16:
+                            bit_depth.dtype = JXL_BIT_DEPTH_FROM_CODESTREAM
                             pixel_format.data_type = JXL_TYPE_UINT16
                             dtype = numpy.uint16
                         else:
@@ -810,6 +818,16 @@ def jpegxl_decode(
                             'JxlDecoderSetImageOutBuffer', status
                         )

+                    if bit_depth.dtype == JXL_BIT_DEPTH_FROM_CODESTREAM:
+                        # do not rescale uint images
+                        status = JxlDecoderSetImageOutBitDepth(
+                            decoder, &bit_depth
+                        )
+                        if status != JXL_DEC_SUCCESS:
+                            raise JpegxlError(
+                                'JxlDecoderSetImageOutBitDepth', status
+                            )
+
                     if is_planar:
                         for channel_index in range(
                             basic_info.num_extra_channels

[jmpfar]

Thanks, looks good! By my very limited testing it creates a 10bit file and the image also renders well