title: "Compressed MP4" abbrev: "cmp4" category: info
docname: draft-lcurley-compressed-mp4-latest submissiontype: IETF # also: "independent", "editorial", "IAB", or "IRTF" number: date: v: 3 area: wit workgroup: moq
fullname: Luke Curley
email: kixelated@gmail.com
normative: RFC9000:
informative: ISOBMFF: title: "Information technology — Coding of audio-visual objects — Part 12: ISO base media file format" target: https://www.iso.org/standard/83102.html date: 2022
--- abstract
Fragmented MP4 (fMP4) is widely used for live streaming, but the ISO Base Media File Format (ISOBMFF) box structure imposes significant per-fragment overhead. Each box header requires 8 bytes (4-byte size + 4-byte type), and payload fields use fixed-width integers (u32/u64). For low-latency streaming with single-frame fragments, this overhead can exceed the media payload itself.
This document defines a compression scheme for ISO BMFF that replaces box headers with compact varint-encoded identifiers and sizes, and defines compressed variants of commonly used boxes with varint payload fields. The scheme reduces per-fragment overhead from ~100 bytes to ~20 bytes while preserving the full box hierarchy.
--- middle
{::boilerplate bcp14-tagged}
Fragmented MP4 (fMP4) is the dominant container format for low-latency live streaming.
Each fragment consists of a moof (movie fragment) box followed by an mdat (media data) box.
The moof box contains metadata describing the media samples in mdat.
For a typical single-frame video fragment, the overhead looks like this:
| Box | Header | Payload | Total |
|---|---|---|---|
| moof | 8 | 0 | 8 |
| mfhd | 8 | 4 | 12 |
| traf | 8 | 0 | 8 |
| tfhd | 8 | 8 | 16 |
| tfdt | 8 | 12 | 20 |
| trun | 8 | 16 | 24 |
| mdat | 8 | 0 | 8 |
| Total | 56 | 40 | 96 |
This 96 bytes of overhead is substantial when a single encoded video frame might only be 100-500 bytes at low bitrates or high frame rates. Audio frames are even smaller, often 4-20 bytes for Opus at low bitrates, making the container overhead several times larger than the payload.
This document defines two layers of compression:
- Header Compression: A compression table (
cmpd) maps varint IDs to box type names. All boxes after themoovuse compressed headers:[varint ID][varint size]instead of[u32 size][4-char type]. - Payload Compression: Compressed variants of common boxes (
cmfh,cfhd,cfdt,crun) replace fixed-width payload fields with varints.
Together, these reduce the per-fragment overhead to approximately 20 bytes.
This document uses the variable-length integer encoding from QUIC {{RFC9000}}, Section 16. The first two bits of the first byte indicate the encoding length:
| 2MSB | Length | Usable Bits | Range |
|---|---|---|---|
| 00 | 1 | 6 | 0-63 |
| 01 | 2 | 14 | 0-16383 |
| 10 | 4 | 30 | 0-1073741823 |
| 11 | 8 | 62 | 0-4611686018427387903 |
In the message formats below, fields marked with (i) use this variable-length integer encoding.
The cmpd box is a standard ISO BMFF box placed inside the moov box.
It defines a mapping from compact varint IDs to 4-character box type names.
cmpd {
Count (i),
Compressed Box Entry (..) ...,
}
Compressed Box Entry {
ID (i),
Name (32),
}
Count: The number of entries in the compression table.
ID: A varint identifier assigned to this box type. IDs SHOULD be assigned starting from 0 to minimize encoding size.
Name: The 4-character ISO BMFF box type name (e.g., moof, mdat, traf).
The cmpd box itself uses a standard ISO BMFF header since it appears inside the moov before compressed encoding takes effect.
A typical compression table for live video streaming:
| ID | Name | Description |
|---|---|---|
| 0 | moof | Movie Fragment |
| 1 | mdat | Media Data |
| 2 | mfhd | Movie Fragment Header |
| 3 | traf | Track Fragment |
| 4 | tfhd | Track Fragment Header |
| 5 | tfdt | Track Fragment Decode Time |
| 6 | trun | Track Run |
With 7 entries using IDs 0-6, each ID fits in a single varint byte. Header compression alone reduces the 56 bytes of box headers (7 boxes x 8 bytes) to 14 bytes (7 boxes x 2 bytes), saving 42 bytes per fragment.
The presence of a cmpd box in the moov signals that all top-level boxes following the moov use compressed box headers.
A standard ISO BMFF box header is:
Standard Box Header {
Size (32),
Type (32),
}
This is replaced with:
Compressed Box Header {
ID (i),
Size (i),
}
ID: The varint identifier from the compression table. The receiver looks up the corresponding 4-character box type name in the cmpd table.
Size: A varint containing the size of the box payload in bytes. Unlike standard ISO BMFF where the size field includes the header itself, the compressed size field contains only the payload length. This avoids the need for extended size fields since varints natively handle large values.
The box hierarchy (nesting) is preserved exactly as in standard ISO BMFF.
Container boxes (e.g., moof, traf) contain nested boxes whose sizes sum to the parent's payload size.
The receiver MUST be able to reconstruct the original uncompressed ISO BMFF structure by reversing the ID-to-name mapping and adjusting size fields.
This section defines compressed variants of commonly used fMP4 boxes. These variants replace fixed-width integer fields with varints, further reducing overhead.
An encoder MAY use the standard box (with a compressed header) OR the compressed variant for any given box. The compression table determines which box type is used.
Replaces mfhd (Movie Fragment Header).
cmfh {
Sequence Number (i),
}
Sequence Number: The fragment sequence number (varint instead of u32).
Standard mfhd uses 4 bytes for the sequence number.
With cmfh, a sequence number under 64 requires only 1 byte.
Replaces tfhd (Track Fragment Header).
cfhd {
Track ID (i),
Flags (i),
Base Data Offset (i), ; present if flags & 0x01
Sample Description Index (i), ; present if flags & 0x02
Default Sample Duration (i), ; present if flags & 0x08
Default Sample Size (i), ; present if flags & 0x10
Default Sample Flags (i), ; present if flags & 0x20
}
Track ID: Identifies the track (varint instead of u32).
Flags: A varint encoding the optional field presence flags. The flag values match the standard tfhd tf_flags semantics but are renumbered for compact varint encoding:
| Flag | Field |
|---|---|
| 0x01 | base-data-offset-present |
| 0x02 | sample-description-index-present |
| 0x08 | default-sample-duration-present |
| 0x10 | default-sample-size-present |
| 0x20 | default-sample-flags-present |
Standard tfhd uses 4 bytes for version/flags and 4 bytes for track ID (minimum 8 bytes).
With cfhd, a single-track stream with no optional fields requires as few as 2 bytes.
Replaces tfdt (Track Fragment Decode Time).
cfdt {
Base Decode Time (i),
}
Base Decode Time: The decode timestamp of the first sample in this fragment (varint instead of u32/u64).
Standard tfdt uses 4 bytes for version/flags plus 4 or 8 bytes for the timestamp (8-12 bytes total).
With cfdt, small timestamps require as few as 1 byte.
Replaces trun (Track Run).
crun {
Sample Count (i),
Flags (i),
Data Offset (i), ; present if flags & 0x01
First Sample Flags (i), ; present if flags & 0x04
Per-Sample Fields (..) ...,
}
Per-Sample Fields {
Sample Duration (i), ; present if flags & 0x100
Sample Size (i), ; present if flags & 0x200
Sample Flags (i), ; present if flags & 0x400
Sample Composition Time Offset (i),; present if flags & 0x800
}
Sample Count: The number of samples in this run (varint instead of u32).
Flags: A varint encoding which optional fields are present. The flag values match the standard trun tr_flags semantics:
| Flag | Field |
|---|---|
| 0x001 | data-offset-present |
| 0x004 | first-sample-flags-present |
| 0x100 | sample-duration-present |
| 0x200 | sample-size-present |
| 0x400 | sample-flags-present |
| 0x800 | sample-composition-time-offset-present |
Standard trun uses 4 bytes for version/flags, 4 bytes for sample count, and 4 bytes per optional field.
With crun, a single-sample run with only sample-size typically requires 4-5 bytes instead of 16.
This section provides a concrete byte-level comparison for a single-frame video fragment.
A typical single-frame fragment with sequence number 42, track ID 1, decode time 3840, and a 200-byte sample:
moof (size=80) 8 bytes
mfhd (size=16) 8 bytes
version=0, flags=0 4 bytes
sequence_number=42 4 bytes
traf (size=56) 8 bytes
tfhd (size=16) 8 bytes
version=0, flags=0x020000 4 bytes
track_id=1 4 bytes
tfdt (size=20) 8 bytes
version=1, flags=0 4 bytes
base_decode_time=3840 8 bytes
trun (size=20) 8 bytes
version=0, flags=0x000200 4 bytes
sample_count=1 4 bytes
sample_size=200 4 bytes
mdat (size=208) 8 bytes
<200 bytes of media data>
Total overhead: 96 bytes (excluding media data).
The same fragment using compressed encoding, with the compression table from the example in Section 4:
moof (id=0, size=11) 2 bytes
cmfh (id=2, size=1) 2 bytes
sequence_number=42 1 byte
traf (id=3, size=6) 2 bytes
cfhd (id=4, size=1) 2 bytes
track_id=1 1 byte
flags=0 0 bytes (no optional fields)
cfdt (id=5, size=2) 2 bytes
base_decode_time=3840 2 bytes
crun (id=6, size=3) 2 bytes
sample_count=1 1 byte
flags=0x200 2 bytes
sample_size=200 2 bytes
mdat (id=1, size=200) 2 bytes
<200 bytes of media data>
Total overhead: ~21 bytes (excluding media data).
This represents a 78% reduction in per-fragment overhead (from 96 bytes to ~21 bytes).
TODO Security
This document registers the following ISO BMFF box types:
| Box Type | Description |
|---|---|
| cmpd | Compression Table |
| cmfh | Compressed Movie Fragment Header |
| cfhd | Compressed Track Fragment Header |
| cfdt | Compressed Track Fragment Decode Time |
| crun | Compressed Track Run |
--- back
{:numbered="false"}
This draft was generated with the assistance of AI (Claude).