avif as a replacement for comic book formats .cbz

[avidseeker]

See https://en.wikipedia.org/wiki/Comic_book_archive. A comic book archive is basically a bunch of .jpg or .png zipped together.

Since avif already supports multi-frame images, would it be a good replacement for .cbz?

i.e: How efficient/optimized would avif be for this task? (e.g: say a comic with 100 pages of high quality scans)

related jpegxl discussion: libjxl/libjxl#4119

[y-guyon]

Thank you for your interest in AVIF.

Since avif already supports multi-frame images, would it be a good replacement for .cbz?

Regarding using AVIF as a good replacement for .cbz for its lossy compression capabilities:
I expect little compression performance gains by inter-frame prediction because the pages of a comic are likely different, meaning not much pixel information can be reused from a page to another. This holds true no matter the image format/codec used.
With that said, I think AVIF would be a good fit for compressing this kind of content at high quality, as long as the source is in very high quality or even raw (otherwise you would carry compression artifacts over to AVIF).

Regarding using AVIF as a good replacement for .cbz because it would all fit in a single file:
You are correct that AVIF supports image sequences, because it is based on HEIF.
However as of today, libavif only supports animations (temporal multi-frame).
So to answer the question, AVIF/HEIF is indeed a good container for storing image sequences. There are metadata tagging options too.

I would also suggest WebP for its good compression performance and tooling support.
However WebP does not support image collections as a container, beyond temporal animations.

---

I tried multiple codecs on four scanned comics pages 1, 2, 3, 4, from the web-images dataset:

Same plot but for decoding time

For the same source image and the same distortion (SSIMULACRA2 ±1.0%),
compared to WebP m6 4:2:0 (quality limited to [70:90]),
images encoded with AVIF s7 4:2:0 are 1.16 times bigger, 1.44 times slower to encode, 1.54 times slower to decode,
images encoded with AVIF s7 4:4:4 are 1.09 times smaller, 1.59 times slower to encode, 1.35 times slower to decode,
images encoded with JPEG XL e7 4:4:4 are 1.05 times bigger, 1.16 times slower to encode, 1.56 times slower to decode.

How to reproduce

wget https://raw.githubusercontent.com/google-research-datasets/web-images/7d8a25e8dbb82069c44e7b1e40d9e1e23a9b1f36/images/fight_comics_0.png
wget https://raw.githubusercontent.com/google-research-datasets/web-images/7d8a25e8dbb82069c44e7b1e40d9e1e23a9b1f36/images/fight_comics_1.png
wget https://raw.githubusercontent.com/google-research-datasets/web-images/7d8a25e8dbb82069c44e7b1e40d9e1e23a9b1f36/images/jumbo_comics_0.png
wget https://raw.githubusercontent.com/google-research-datasets/web-images/7d8a25e8dbb82069c44e7b1e40d9e1e23a9b1f36/images/jumbo_comics_1.png

git clone -b v0.5.1 --depth 1 https://github.com/webmproject/codec-compare-gen.git
cd codec-compare-gen
./deps.sh
cmake -S . -B build -DCMAKE_CXX_COMPILER=clang++
cmake --build build --parallel
mkdir ../json
mkdir ../encoded
build/ccgen \
  --codec webp 420 0 --codec webp 420 1 --codec webp 420 2 --codec webp 420 3 --codec webp 420 4 --codec webp 420 5 --codec webp 420 6 \
  --codec avif 420 0 --codec avif 420 1 --codec avif 420 2 --codec avif 420 3 --codec avif 420 4 --codec avif 420 5 --codec avif 420 6 --codec avif 420 7 --codec avif 420 8 --codec avif 420 9 \
  --codec avif 444 0 --codec avif 444 1 --codec avif 444 2 --codec avif 444 3 --codec avif 444 4 --codec avif 444 5 --codec avif 444 6 --codec avif 444 7 --codec avif 444 8 --codec avif 444 9 \
  --codec jpegxl 444 1 --codec jpegxl 444 2 --codec jpegxl 444 3 --codec jpegxl 444 4 --codec jpegxl 444 5 --codec jpegxl 444 6 --codec jpegxl 444 7 --codec jpegxl 444 8 --codec jpegxl 444 9 --codec jpegxl 444 10 \
  --codec jpegmoz 444 --codec jpegmoz 420 \
  --lossy --threads 127 --results_folder ../json/ --encoded_folder ../encoded/ --metric_binary_folder third_party/ \
  -- ../*.png

git clone -b v0.5.4 --depth 1 https://github.com/webmproject/codec-compare.git
cd codec-compare
# Modify default_batches.json to point to the JSON files generated by ccgen in the "json" folder
npm i
npm run dev

The tested corpus is tiny but it gives some idea of what to expect.
I used WebP at maximum effort as a baseline, at high quality range 70 to 90. Other codec efforts are somewhat similar in encoding timings.
SSIMULACRA2 is used as an objective visual quality evaluation metric for convenience but one should manually inspect the decoded images to make sure the encoding quality setting is optimal for each codec.