Format Comparison
======================================

Overview
------------------------------

.. list-table::
    :header-rows: 1
    :align: left

    * - Format
      - Lossy
      - Lossless
      - Compatibility
      - Transparency
    * - JPEG XL
      - High
      - High
      - Low
      - Yes
    * - AVIF
      - High
      - \-
      - Medium
      - Yes
    * - WebP
      - Medium
      - Medium
      - Medium-High
      - Yes
    * - JPEG
      - Low
      - \-
      - High
      - No
    * - PNG
      - \-
      - Low
      - High
      - Yes

JPEG XL vs AVIF
------------------------------

Brief Description
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

JPEG XL and AVIF are both modern image formats offering high compression ratios. They outperform HEIC, WebP, and JPEG by a wide margin.

Introduction
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Before evaluating lossy efficiency, it's important to understand the difference between photographic and synthetic datasets.

* **Synthetic (non-photographic)**: images created entirely on a computer without using a camera. 
* **Photographic**: images captured using a camera.
* **Mixed**: datasets featuring both synthetic nad photographic images. Hybrid images featuring both photographic and synthetic elements qualify as well.

CID22
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Provided by Cloudinary (JPEG XL developers). CID22 is a mixed dataset, featuring both photographic, and synthetic images. `Source <https://cloudinary.com/labs/cid22>`_

JPEG XL performs well on this mixed dataset.

.. figure:: _static/cid22_all.png

    CID22 - lossy performance on the entire dataset.

JPEG XL performs best when tested on photographic data.

.. figure:: _static/cid22_landscape_nature.png

    CID22 - lossy performance on landscape and nature.

When isolated to diagrams and charts (synthetic images), AVIF (AOM AV1) outperforms JPEG XL.

.. figure:: _static/cid22_diagram_chart.png

    CID22 - lossy performance on diagram and charts.

Try this `interactive comparison <https://cloudinary.com/labs/cid22/plots>`_.

This dataset was covered in more detail in the official JPEG XL presentation by Jon Sneyers `here <https://docs.google.com/presentation/d/1LlmUR0Uoh4dgT3DjanLjhlXrk_5W2nJBDqDAMbhe8v8/edit#slide=id.g2c51e8f95e1_0_422>`_.

Personally Observed Performance
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

This is a subjective overview. Efficiency on your dataset may vary.

**Lossy:**

.. list-table::
    :header-rows: 1
    :align: left
    
    * - Dataset Type
      - Most Efficient
      - Dataset Example
      - Explanation
    * - Photographic
      - JPEG XL
      - Photographs, scans.
      - JPEG XL excels at compressing photographic material thanks to the VarDCT mode.
    * - Synthetic (non-photographic)
      - \*AVIF (AOM AV1)
      - Screenshots, diagrams, vector graphics, Anime, digital art.
      - AOM AV1 has better efficiency on synthetic images than JPEG XL. \*Occasionally, JPEG XL can come out ahead, but this isn't as common, and happens only with very complex or large images.
    * - Mixed
      - JPEG XL
      - A mix of photographic and non-photographic images.
      - JPEG XL is better for unsorted datasets because how much more efficient it is on photographic images.

**Lossless:**

JPEG XL lossless mode offers the `best <https://codepoems.eu/posts/how-to-reduce-the-size-of-lossless-images/>`_ efficiency overall. WebP may perform better on certain images, but JPEG XL is the best on average. AVIF lossless mode is oftentimes worse than PNG.

Transcoding Speed
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

.. list-table::
    :header-rows: 1
    :align: left

    * - Format
      - Encoder
      - Speed
    * - JPEG
      - JPEGLI
      - 10.26X
    * - JPEG
      - libjpeg
      - 1.09x
    * - JPEG XL
      - libjxl
      - 1x
    * - AVIF
      - AOM AV1
      - 0.7x

.. raw:: html
    
    <small>Parameters used: Speed / Effort: default. Input: PNG.</small>

Capabilities
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

.. list-table::
    :header-rows: 1
    :align: left

    * - Feature
      - JPEG XL
      - AVIF
    * - Max Resolution
      - 1,073,741,823 x 1,073,741,823
      - \*65,536 x 65,536
    * - Max Bit Depth
      - 32 bit
      - 12 bit
    * - Compatibility
      - Low
      - Medium
    * - Lossy Efficiency
      - Excellent
      - Excellent
    * - Lossless Efficiency
      - Excellent
      - \-

\* Only 8193 x 4320 natively, larger sizes are achieved with tiling (which lowers efficiency and introduces artifacts).

JPEG XL has extra features, such as:

* Lossless JPEG Transcoding
* (Better) Progressive Decoding
* Generation Loss Resilience
* Photon Noise
* and more...