PNG of 8 x 8 pixels: 488 bytes, signature 89 50 4e 47 0d 0a 1a 0a

type  data bytes  crc ok  kind        what
IHDR  13          True    critical    width 8, height 8, bit depth 8, color type 2
tEXt  182         True    ancillary   keyword prompt
tEXt  117         True    ancillary   keyword workflow
IDAT  108         True    critical    zlib data, 200 bytes once unpacked
IEND  0           True    critical    end marker

the last 12 bytes of every PNG: 00 00 00 00 49 45 4e 44 ae 42 60 82   crc of the bytes IEND: ae426082
json.dumps writes only ASCII by default: "snowman \u2603"   the raw character does not fit Latin-1: True
one bit flipped inside the prompt text: chunks failing their crc: ['tEXt prompt']
only the critical chunks kept: 488 -> 165 bytes, text chunks left: 0, same pixels: True

WAV: 32044 bytes = 44 header bytes + 32000 data bytes; RIFF size field 32036 = file size - 8
fmt: format 1 (PCM), 2 channels, 8000 Hz, byte rate 32000 = 8000 x 2 x 2, block align 4, 16 bits per sample
first two sample pairs as integers (0, 0, 4064, 5945), as floats (0.0, 0.0, 0.124, 0.1814), extremes -1.0 and 0.99997
the AUDIO tensor ComfyUI would hold for this file: waveform [1, 2, 8000], sample_rate 8000
with an odd-sized 'note' chunk before the data: data starts at byte 56, not 44; walking the chunks finds True, reading at 44 finds False

index last : ftyp at 0 (24 bytes), mdat at 24 (1000008 bytes), moov at 1000032 (208 bytes)
   bytes to read before the index (moov) is complete: 1,000,240 of 1,000,240

index first: ftyp at 0 (24 bytes), moov at 24 (208 bytes), mdat at 232 (1000008 bytes)
   bytes to read before the index (moov) is complete: 232 of 1,000,240

box with size field 1: real size read from the next 8 bytes = 21; box with size field 0 in a 12-byte stream: size 12
