webp/doc/webp-container-spec.txt

<!--

Although you may be viewing an alternate representation, this document
is sourced in Markdown, a light-duty markup scheme, and is optimized for
the [kramdown](http://kramdown.rubyforge.org/) transformer.

See the accompanying README. External link targets are referenced at the
end of this file.

-->


WebP Container Specification
============================

_Working Draft, v0.1, 20111004_


* TOC placeholder
{:toc}


Introduction
------------

WebP is a still image format that uses the VP8 key frame encoding, and
possibly other encodings in the future, to compress image data in a
lossy way. The VP8 encoding should make it more efficient than currently
used formats. It is optimized for fast image transfer over the network
(e.g., for websites). However, it also aims for feature parity (like
Color Profile, XMP Metadata, Animation, etc.) with other formats. This
document describes the structure of a WebP file.

The first version of WebP handled only the basic use case:  a file
containing a single image (being one VP8 key frame), with no metadata.
The use of a RIFF container permits additional feature support.  This
document describes additional support for:

  * **Metadata and color profiles.** We specify chunks that can contain
    this information, as other popular formats do.

  * **Tiling.** A single VP8 frame has an inherent limitation for width
    or height of 2^14 pixels, and a 512kB limit on the size of the first
    compressed partition. To support larger images, we support images
    that are composed of multiple tiles, each encoded as a separate VP8
    frame. All tiles form logically a single image:  they have common
    metadata, color profile, etc. Tiling may also improve efficiency for
    larger images, e.g., grass can be encoded differently than sky.

  * **Animation.** An image may have pauses between frames, making it
    an animation.

  * **Transparency.** An image may have transparency, i.e., an alpha channel for
    each frame/tile.

Files not using these new features are backward compatible with the
original format. Use of these features will produce files that are not
compatible with older programs.


Terminology &amp; Basics
------------------------

A WebP file contains either a still image (i.e., an encoded matrix of
pixels) or an animation (see below), with possibly a color profile,
metadata, etc. In case we need to refer only to the matrix of pixels,
we will call it the **_canvas_** of the image.

The canvas of an image is built from one or multiple tiles. Each tile
is a separately encoded VP8 key frame (other encodings are possible in
the future). Building an image from several tiles allows us to overcome
the size limitations of a single VP8 frame. Tiles are an internal detail
of the file:  they are not supposed to be exposed to the user.

Below are additional terms used throughout this document:

Code that reads WebP files is referred to as a **_reader_**, while
code that writes them is referred to as a **_writer_**.

A 16-bit, little-endian, unsigned integer will be denoted as
**_uint16_**.

A 32-bit, little-endian, unsigned integer will be denoted as
**_uint32_**.

The basic element of a RIFF file is a **_chunk_**. It consists of:

  * 4 ASCII characters that will be called the **_chunk tag_**.

  * uint32 with the size of the chunk content (that will be denoted as
    **_ckSize_**).

  * _ckSize_ bytes of content.

  * If _ckSize_ is odd, a single padding byte that **SHOULD** be `0`.

A chunk with a tag "ABCD" will be also called a **_chunk of type_**
"ABCD". Note that, in this specification, all chunk tag characters are
in file order, not in byte order of a uint32 of any particular
architecture.

Note that the padding **MUST** be added to the last chunk of the file.

A **_list of chunks_** is a concatenation of multiple chunks. We will
refer to the first chunk as having _position_ 0, the second as position
1, etc. By _chunk with index 0 among "ABCD"_ we mean the first chunk
among the chunks of type "ABCD" in the list, the _chunk with index 1
among "ABCD"_ is the second such chunk, etc.

A WebP file **MUST** begin with a single chunk with a tag "RIFF". All
other defined chunks are contained within this chunk. The file **SHOULD
NOT** contain anything after it.

The maximum size of RIFF's _ckSize_ is 2^32 minus 10 bytes.  The size
of the whole file is at most 4GiB minus 2 bytes.

**Note:** some RIFF libraries are said to have bugs when handling files
larger than 1GiB or 2GiB. If you are using an existing library, check
that it handles large files correctly.

The first four bytes of the RIFF chunk contents (i.e., bytes 8-11 of the
file) **MUST** be the ASCII string "WEBP". They are followed by a list
of chunks. Note that as the size of any chunk is even, the size of the
RIFF chunk is also even.

The contents of the chunks in that list will be described in the
following sections.

**Note:** RIFF has a convention that all-uppercase chunks are standard
chunks that apply to any RIFF file format, while chunks specific to a
file format are all-lowercase. WebP doesn't follow this convention.


Single-image WebP Files
-----------------------

First, we will describe a subset of WebP files:  files containing only
one image. Later, we will define multi-image files, which contain
several images.


### Chunks Layout

This section describes which chunks may appear in a single-image WebP
file, and their order. The contents of these chunks will be described
in subsequent sections.

The first chunk inside the RIFF chunk **MUST** have a tag of "VP8 "
(note that the fourth character is a space, and is significant) or
"VP8X". Other tags for the first chunk **MAY** be introduced by future
specifications if new encodings are added. This tag of the first chunk
determines which of the two possible layouts is used.

**Rationale:**  We fix the possible tags of the first chunk so that it
is possible to introduce other codecs, to keep the "WEBP" signature at
the beginning of the RIFF chunk while still being able to check the
codec used by the image by inspecting the byte stream at a fixed
position.

The two possible layouts will be called _images without special layout_
and _images with special layout_.


#### Images Without Special Layout

If the first subchunk of RIFF has the tag "VP8 ", the file contains an
_image without special layout_.

This layout **SHOULD** be used if the image doesn't require advanced
features: color profiles, XMP metadata, animation or tiling. Files with
this layout are smaller and supported by older software.

Such images consist of:

  * A "VP8 " chunk with the bitstream of the single tile.

**Example:** An example layout of such a file is as follows:

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
RIFF/WEBP
+- VP8 (bitstream of the single tile of the image)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~


#### Images With Special Layout

If the first subchunk of RIFF has the tag "VP8X", the file contains an
_image with special layout_.

**Note:**  Older readers may not support images with special layout.

Such an image consists of:

  * A "VP8X" chunk with information about features used in the file.

  * An optional "ICCP" chunk with color profile.

  * An optional "LOOP" chunk with animation control data.

  * Data for all the frames.

  * An optional "META" chunk with XMP metadata.

  * Some other chunk types may be defined by future specifications and
    placed anywhere in the file.

As will be described in the "VP8X" chunk description, by checking a
flag one can distinguish animated and non-animated images. A
non-animated image has exactly one frame. An animated one may have
multiple frames. Data for each frame consists of:

  * An optional "FRM " (fourth character is a significant space) chunk
    with animation frame metadata. It **MUST** be present in animated
    images at the beginning of data for that frame. It **MUST NOT** be
    present in non-animated images.

  * An optional "TILE" chunk with tile position metadata. It **MUST** be
    present at the beginning of data for an image that's represented as
    multiple tile images.

  * An optional "ALPH" chunk with alpha bitstream of the tile. It **MUST** be
    present for an image containing transparency. It **MUST NOT** be present
    in non-transparent images.

  * A "VP8 " chunk with the bitstream of the tile.

All chunks **MUST** be placed in the same order as listed above (except
for unknown chunks, which **MAY** appear anywhere). If a chunk appears
in the wrong place, the file is invalid, but readers **MAY** parse the
file, ignoring the chunks that come too late.

**Rationale:**  Setting the order of chunks should allow quicker file
parsing. For example, if an ICCP chunk does not appear in its required
position, a decoder can choose to stop searching for it.  The rule of
ignoring late chunks should make programs that need to do a full search
give the same results as the ones stopping early.

**Example:** An example layout of a non-animated, tiled image without
transparency may look as follows:

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
RIFF/WEBP
+- VP8X (descriptions of features used)
+- ICCP (color profile)
+- TILE (First tile parameters)
+- VP8 (bitstream - first tile)
+- TILE (Second tile parameters)
+- VP8 (bitstream - second tile)
+- TILE (third tile parameters)
+- VP8 (bitstream - third tile)
+- TILE (fourth tile parameters)
+- VP8 (bitstream - fourth tile)
+- META (XMP metadata)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

**Example:** An example layout of an animated image with transparency may look
as follows:

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
RIFF/WEBP
+- VP8X (descriptions of features used)
+- LOOP (animation control parameters)
+- FRM (first animation frame parameters)
+- ALPH (alpha bitstream - first image frame)
+- VP8 (bitstream - first image frame)
+- FRM (second animation frame parameters)
+- ALPH (alpha bitstream - second image frame)
+- VP8 (bitstream - second image frame)
+- META (XMP metadata)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~


### Assembling the Canvas from Tiles and Animation

Contents of the chunks will be described in subsequent sections. Here we
provide an overview of how they are used to assemble the canvas. The
notation _VP8X.canvasWidth_ means the field in the "VP8X"
described as _canvasWidth_.

Decoding a non-animated canvas **MUST** be equivalent to the following
pseudocode:

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
assert not VP8X.flags.haveAnimation
canvas ← new black image of size
         VP8X.canvasWidth x VP8X.canvasHeight.
tile_params.tileCanvasX = tile_params.tileCanvasY = 0
for chunk in data_for_all_frames:
    if chunk.tag == "TILE":
        assert No other TILE chunk after the last "VP8 " chunk
        assert No ALPH chunk after the last "VP8 " chunk
        tile_params = chunk
    if chunk.tag == "ALPH":
        assert No other ALPH chunk after the last "VP8 " chunk
        tile_params.alpha = alpha_data
    if chunk.tag == "VP8 ":
        render image in chunk in canvas with top-left corner in
            (tile_params.tileCanvasX, tile_params.tileCanvasY).
        tile_params.tileCanvasX = tile_params.tileCanvasY = 0
    Ignore unknown chunks
canvas contains the decoded canvas.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Decoding an animated canvas **MUST** be equivalent to the following
pseudocode:

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
assert VP8X.flags.haveAnimation
canvas ← new black image of size
         VP8X.canvasWidth x VP8X.canvasHeight.
if LOOP.loopCount==0:
    LOOP.loopCount=∞
current_FRM ← nil
for LOOP.loop = 0, ..., LOOP.loopCount-1
    assert First chunk in data_for_all_frames is a FRM
    for chunk in data_for_all_frames:
        if chunk.tag == "FRM ":
            assert No other FRM chunk after the last "VP8 " chunk
            assert No ALPH chunk after the last "VP8 " chunk
            if current_FRM != nil:
                Show the contents of canvas for
                    current_FRM.frameDuration * 1ms.
            current_FRM = chunk
        if chunk.tag == "ALPH":
            assert No other ALPH chunk after the last "VP8 " chunk
            tile_params.alpha = alpha_data
        if chunk.tag == "VP8 ":
            assert tile_params.tileCanvasX >= current_FRM.frameX
            assert tile_params.tileCanvasY >= current_FRM.frameY
            assert tile_params.tileCanvasX + chunk.tileWidth >=
                current_FRM.frameX + current_FRM.frameWidth
            assert tile_params.tileCanvasY + chunk.tileHeight >=
                current_FRM.frameX + current_FRM.frameHeight
            render image in chunk in canvas with top-left corner in
                (tile_params.tileCanvasX, tile_params.tileCanvasY).
            tile_params.tileCanvasX = tile_params.tileCanvasY = 0
        Ignore unknown chunks
canvas contains the decoded canvas.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

As described earlier, if an assert related to chunk ordering fails, the
reader **MAY** ignore the badly-ordered chunks instead of failing to
decode the file.


### Bitstream Chunks (VP8)

These chunks contain compressed image data. Currently, the only allowed
bitstream is VP8, using "VP8 " (note the significant fourth-character
space) as its tag. We will refer to all chunks with this tag as
**_bitstream chunks_**. As described earlier, images without special
layout have a single bitstream chunk as the first subchunk of RIFF,
while images with special layout may contain several of them, one for
each tile.

The content of a "VP8 " chunk **MUST** be one VP8 key frame (with
optional padding.  See below).

The current [VP8 Data Format and Decoding Guide][vp8spec] can be found
at the IETF website, <http://www.ietf.org/>. Note that the VP8 frame
header contains the VP8 frame width and height. That is assumed to be
the width and height of the tile.

The VP8 specification describes how to decode the image into Y'CbCr
format. To convert to RGB, Rec. 601 **SHOULD** be used.

For compatibility with older readers, if the size of the frame is odd,
writers **SHOULD** append a padding byte (preferably `0`) inside the
chunk contents, making the chunk's _ckSize_ even. Newer readers
**MUST** support odd-sized bitstream chunks.


### VP8X Chunk (Special Layout)

As described earlier, a chunk with tag "VP8X", is the first chunk of
images with special layout. It is used to enable advanced features of
WebP.

The content of the chunk is as follows:

  * **uint32** flags. The following bits are currently used (with `0`
    being the least significant bit):

    * bit 0: _hasTile_:  Set if the image is represented by Tiles.

    * bit 1: _hasAnimation_:  Set if the file is an animation. Data in
      "LOOP" and "FRM " chunks should be used to control the animation.

    * bit 2: _hasIccp_:  Set if the file contains an "ICCP" chunk with a
      color profile. If a file contains an "ICCP" chunk but this bit is
      not set, the error is flagged while constructing the
      Mux-Container.

    * bit 3: _hasMetadata_:  Set if the file contains a "META" chunk
      with a XMP metadata. If a file contains an "META" chunk but this
      bit is not set, the error is flagged while constructing the
      Mux-Container.

Future specifications **MAY** define other bits in flags. Bits not
defined by this specification **MUST** be preserved when modifying the
file.

  * **uint32** _canvasWidth_:  Width of the canvas in pixels (after the
    optional rotation or symmetry; see below).

  * **uint32** _canvasHeight_:  Height of the canvas in pixels (after
    the optional rotation or symmetry; see below).

Future specifications **MAY** add more fields. If a chunk of larger size
is found, programs **MUST** ignore the extra bytes but **MUST** preserve
them when modifying the file.


### LOOP Chunk (Global Animation Parameters)

For images that are animations, this chunk contains the global
parameters of the animation.

This chunk **MUST** appear if the _hasAnimation_ flag in chunk VP8X is
set. If the _hasAnimation_ flag is not set and this chunk is present,
it **MUST** be ignored.

The content of the chunk is as follows:

  * **uint16** _loopCount_:  For animations, the number of times to loop
    the animation. `0` means infinitely.

Future specifications **MAY** add more fields. If a chunk of larger
size is found, programs **MUST** ignore the extra bytes but **MUST**
preserve them when modifying the file.


### FRM Chunk (Per-frame Animation Parameters)

For images that are animations, these chunks contain the per-frame
parameters of the animation.

The content of the chunk is as follows:

  * **uint32** _frameX_:  X coordinate of the upper left corner of the
    frame. For images using the VP8 codec, this value **MUST** be
    divisible by `32`. Other codecs **MAY** specify other constraints.
    Described in more detail later.

  * **uint32** _frameY_:  Y coordinate of the upper left corner of the
    frame. For images using the VP8 codec, this value **MUST** be
    divisible by `32`. Other codecs **MAY** specify other constraints.
    Described in more detail later.

  * **uint32** _frameWidth_:  Width of the frame. For images using the
    VP8 codec, this value **MUST** be divisible by `16`, or be such that
    _frameX + frameWidth == canvasWidth_. Other codecs **MAY** specify
    other constraints. Described in more detail later.

  * **uint32** _frameHeight_:  Height. For images using the VP8 codec,
    this value **MUST** be divisible by `16`, or be such that _frameY +
    frameHeight == canvasHeight_. Other codecs **MAY** specify other
    constraints. Described in more detail later.

  * **uint16** _frameDuration_:  Time to wait before displaying the next
    tile, in 1ms units.

**Rationale:**  The requirement for corner coordinates to be divisible
by `32` means that pixels on U and V planes are aligned to a 16-byte
boundary (even after a rotation), which may help with vector
instructions on some architectures. This makes the tiles also align to
16-pixel macroblock boundaries.

**Rationale:**  The requirement for the width and height to be
divisible by `16` or touching the edge of the canvas simplifies the
handling of macroblocks that are on the edge of a tile.  VP8 decoders
can overwrite pixels outside the boundary in such a macroblock, and this
guarantees they won't overwrite any data.

Future specifications **MAY** add more fields. If a chunk of larger
size is found, programs **MUST** ignore the extra bytes but **MUST**
preserve them when modifying the file.


### TILE Chunks (Tile Parameters)

This chunk contains information about a single tile and describes the
bitstream chunk that follows it.

The contents of such a chunk are as follows:

  * **uint32** _tileCanvasX_:  X coordinate of the upper left corner of
    the tile. For VP8 tiles, this value **MUST** be divisible by `32`.
    Other codecs **MAY** specify other constraints.

  * **uint32** _tileCanvasY_:  Y coordinate of the upper left corner of
    the tile. For VP8 tiles, this value **MUST** be divisible by `32`.
    Other codecs **MAY** specify other constraints.

Future specifications **MAY** add more fields. If a chunk of larger size
is found, programs **MUST** ignore the extra bytes but **MUST** preserve
them when modifying the file.

As described earlier, the TILE chunk is followed by VP8 data. From that
chunk we can read the height and width of the tile.  These we denote as
_tileWidth_ and _tileHeight_. In the case of VP8, we have the following
constraints:

  * The width of a tile **MUST** be divisible by `16`, or _tileCanvasX +
    tileWidth == canvasWidth_ **MUST** be true.

  * The height of a tile **MUST** be divisible by `16`, or
    _tileCanvasY + tileHeight == canvasHeight_ **MUST** be true.


### ALPH Chunks (Alpha Bitstreams)

This optional chunk contains encoded alpha data for a single tile. Either
**ALL or NONE** of the tiles must contain this chunk.

The alpha channel can be encoded either losslessly or with lossy preprocessing
(quantization). After the optional preprocessing, the alpha values are encoded
with a lossless compression method like zlib. Work is in progress to improve the
compression gain further by exploring alternate compression methods and hence,
the bit-stream for the Alpha-chunk is still experimental and expected to change.

The contents of such a chunk are as follows:

  * Byte 0 lower nibble: The _compression method_ used. Currently two methods
    are supported:

    * 0 --> No compression

    * 1 --> Backward reference counts encoded with arithmetic encoder.

  * Byte 0 upper nibble: The _filtering method_ used. Currently the following
    methods are supported:

    * 0 --> No filter

    * 1 --> Horizontal filter

    * 2 --> Vertical filter

    * 3 --> Gradient filter

  * Byte 1:  _Reserved_. **Should** be 0.

  * Byte 2 onwards:  _Encoded alpha bitstream_.


### ICCP Chunk (Color Profile)

An optional "ICCP" chunk contains an ICC profile. There **SHOULD** be
at most one such chunk. The first byte of the chunk is the compression
type. Two values are currently defined:  a value of `0` means no
compression, while a value of `1` means deflate/inflate compression. It
is followed by a compressed or non-compressed ICC profile.  See
<http://www.color.org> for specifications.

The color profile can be a v2 or v4 profile. If this chunk is missing,
sRGB **SHOULD** be assumed.


### META Chunk (Compressed XMP Metadata)

Such a chunk (if present) contains XMP metadata. There **SHOULD** be at
most one such chunk. If there are more such chunks, readers **SHOULD**
ignore all except the first one. The first byte specifies compression
type. Two values are currently defined:  a value of `0` means no
compression, while a value of `1` means deflate/inflate compression. It
is followed by a compressed or non-compressed XMP metadata packet.

XMP packets are XML text as specified in the [XMP Specification Part
1][xmpspec]. The chunk tag is different from the one specified by Adobe
for WAV and AVI (also RIFF formats), because we have the option of
compression.

Additional guidance about handling metadata can be found in the
Metadata Working Group's [Guidelines for Handling Metadata][metadata].
Note that the sections of the document about reconciliation of EXIF,
XMP and IPTC-IIM don't apply to WebP.  As WebP supports only XMP, no
reconciliation is necessary.


### Other Chunks

A file **MAY** contain other chunks, defined in some future
specification. Such chunks **MUST** be ignored, but preserved. Writers
**SHOULD** try to preserve them in their original order.


[vp8spec]:  http://tools.ietf.org/html/rfc6386
[xmpspec]:  http://www.adobe.com/content/dam/Adobe/en/devnet/xmp/pdfs/XMPSpecificationPart1.pdf
[metadata]: http://www.metadataworkinggroup.org/pdf/mwg_guidance.pdf