[Question] How optimally load sequence of 16 bit images?

Hi, I'am seeing for optimal way of loading stack of PNG uint16 images with fixed sequence length 1 < N <= 8 and GT image as label.
I thinking about few cases:

1. N images saved in dir and GT in other dir
2. Save all images as tfrecord with corresponding labels for stack and GT

For 1 case it looks like SequenceReader should be used, but I am not see an option to load GT image - should I just load a sequence with GT and take a split from this or maybe another way?

For 2 case I met the next problem - I can customize TFRecord reader, but can't decode that due to stack of images, what the right way to customly decode this sequence?
Currently I use next code based from tensorflow RecordDataset and want to adapt this with DALI:

labeled_tfrec_format = {
'n_imgs': tf.io.FixedLenFeature([], tf.int64),
'height': tf.io.FixedLenFeature([], tf.int64),
'width': tf.io.FixedLenFeature([], tf.int64),
'channels': tf.io.FixedLenFeature([], tf.int64),
'image_raw': tf.io.FixedLenFeature([], tf.string),
'gt_raw': tf.io.FixedLenFeature([], tf.string)}
def read_tfrecord(example):
        example = tf.io.parse_single_example(example, labeled_tfrec_format)
        image = tf.io.decode_raw(example['image_raw'], tf.uint16)
        image = tf.reshape(image, (example['N'], example['H'], example['W'], example['channels']))
        gt = tf.io.decode_raw(example['gt_raw'], tf.uint16)
        gt = tf.reshape(gt, (example['H'], example['W'], example['channels']))
        return image, gt
ds = tf.data.TFRecordDataset(some_pathes)
ds = ds.map(read_tfrecord)
ds = ds.map(..)  # some reshapes, crops, etc...

Start with the SequenceReader and TFRecord reader paths mentioned in the issue, then compare them with the provided Python parsing example. Determine whether loading a variable-length image sequence with a separate GT image is already supported or needs a documented capability; done means a clear recommended approach for both storage cases and decoding uint16 stacks.