FFV1 Video Codec Specification

by Michael Niedermayer <michaelni@gmx.at>

Table of Contents

1 Introduction

The FFV1 video codec is a simple and efficient lossless intra-frame only codec.
The latest version of this document is available at https://raw.github.com/FFmpeg/FFV1/master/ffv1.lyx
This document assumes familiarity with mathematical and coding concepts such as Range coding and YCbCr colorspaces.

2 Terms and Definitions

2.1 Terms

The key words "MUST", "MUST NOT", "SHOULD", and "SHOULD NOT" in this document are to be interpreted as described in RFC 2119 [1].
For reference, below is an excerpt of RFC 2119:
“MUST” means that the definition is an absolute requirement of the specification.
“MUST NOT” means that the definition is an absolute prohibition of the specification.
“SHOULD” means that there may exist valid reasons in particular circumstances to ignore a particular item, but the full implications must be understood and carefully weighed before choosing a different course.
“SHOULD NOT” means that there may exist valid reasons in particular circumstances when the particular behavior is acceptable or even useful, but the full implications should be understood and the case carefully weighed before implementing any behavior described with this label.

2.2 Definitions

ESC Escape symbol to indicate that the symbol to be stored is too large for normal storage and a different method is used to store it.
MSB Most significant bit, the bit that can cause the largest change in magnitude of the symbol
RCT Reversible Color Transform, a near linear, exactly reversible integer transform that converts between RGB and YCbCr representations of a sample.
VLC Variable length code

3 General Description

Each frame is split in 1 to 4 planes (Y, Cb, Cr, Alpha). In the case of the normal YCbCr colorspace the Y plane is coded first followed by the Cb and Cr planes, if an Alpha/transparency plane exists, it is coded last. In the case of the JPEG2000-RCT colorspace the lines are interleaved to improve caching efficiency since it is most likely that the RCT will immediately be converted to RGB during decoding; the interleaved coding order is also Y,Cb,Cr,Alpha.
Samples within a plane are coded in raster scan order (left->right, top->bottom). Each sample is predicted by the median predictor from samples in the same plane and the difference is stored see3.6↓.

3.1 Border

For the purpose of the predictior and context, samples above the coded slice are assumed to be 0; samples to the right of the coded slice are identical to the closest left sample; samples to the left of the coded slice are identical to the top right sample (if there is one), otherwise 0.
0 0 0 0 0 0
0 0 0 0 0 0
0 0 a b c c
0 a d e e
0 d f g h h

3.2 Median predictor

median(left, top, left + top - diag)
left, top, diag are the left, top and left-top samples
Note, this is also used in JPEG-LS and HuffYuv[1, 3].

3.3 Context

tl t tr
L l X
The quantized sample differences L-l, l-tl, tl-t, t-T, t-tr are used as context:
context = Q0[l − tl] + |Q0|(Q1[tl − t] + |Q1|(Q2[t − tr] + |Q2|(Q3[L − l] + |Q3|Q4[T − t])))
If the context is smaller than 0 then -context is used and the difference between the sample and its predicted value is encoded with a flipped sign.

3.4 Quantization

There are 5 quantization tables for the 5 sample differences, both the number of quantization steps and their distribution are stored in the bitstream. Each quantization table has exactly 256 entries, and the 8 least significant bits of the sample difference are used as index:
Qi[a − b] = Tablei[(a − b)&255]

3.5 Colorspace

3.5.1 JPEG2000-RCT

Cb = b − g
Cr = r − g
Y = g + (Cb + Cr) >  > 2
g = Y − (Cb + Cr) >  > 2
r = Cr + g
b = Cb + g

3.6 Coding of the sample difference

Instead of coding the n+1 bits of the sample difference with Huffman-, or Range coding (or n+2 bits, in the case of RCT), only the n (or n+1) least significant bits are used, since this is sufficient to recover the original sample. In the equation below, the term “bits” represents bits_per_raw_sample+1 for RCT or bits_per_raw_sample otherwise:
coder_input = [(sample_difference + 2bits − 1)&(2bits − 1)] − 2bits − 1

3.6.1 Range coding mode

Early experimental versions of FFV1 used the CABAC Arithmetic coder from H.264[2], but due to the uncertain patent/royality situation, as well as its slightly worse performance, CABAC was replaced by a range coder based on an algorithm defined by G. Nigel N. Martin in 1979 [6].
Binary values
To encode binary digits efficiently a range coder is used. Ci is the i-th Context. Bi is the i-th byte of the bytestream. bi is the i-th range coded binary value, S0, i is the i-th initial state, which is 128. The length of the bytestream encoding n binary symbols is jn bytes.
ri = (RiSi, Ci)/(28)
Si + 1, Ci = zero_stateSi, Ci li = Li ti = Ri − ri bi = 0 Li < Ri − ri Si + 1, Ci = one_stateSi, Ci li = Li − Ri + ri ti = ri bi = 1 Li ≥ Ri − ri
Si + 1, k = Si, k Ci ≠ k
Ri + 1 = 28ti Li + 1 = 28li + Bji ji + 1 = ji + 1 ti < 28 Ri + 1 = ti Li + 1 = li ji + 1 = ji ti ≥ 28
R0 = 65280
L0 = 28B0 + B1
j0 = 2
Non binary values
To encode scalar integers it would be possible to encode each bit separately and use the past bits as context. However that would mean 255 contexts per 8-bit symbol which is not only a waste of memory but also requires more past data to reach a reasonably good estimate of the probabilities. Alternatively assuming a Laplacian distribution and only dealing with its variance and mean (as in Huffman coding) would also be possible, however, for maximum flexibility and simplicity, the chosen method uses a single symbol to encode if a number is 0 and if not encodes the number using its exponent, mantissa and sign. The exact contexts used are best described by the following code, followed by some comments.
void put_symbol(RangeCoder *c, uint8_t *state, int v, int is_signed) {
    int i;
    put_rac(c, state+0, !v);
    if (v) {
        int a= ABS(v);
        int e= log2(a);
        for (i=0; i<e; i++)
            put_rac(c, state+1+MIN(i,9), 1);  //1..10

        put_rac(c, state+1+MIN(i,9), 0);
        for (i=e-1; i>=0; i--)
            put_rac(c, state+22+MIN(i,9), (a>>i)&1); //22..31

        if (is_signed)
            put_rac(c, state+11 + MIN(e, 10), v < 0); //11..21
Initial values for the context model
At keyframes all range coder state variables are set to their initial state.
State transition table
one_statei = default_state_transitioni + state_transition_deltai
zero_statei = 256 − one_state256 − i
  0,  0,  0,  0,  0,  0,  0,  0, 20, 21, 22, 23, 24, 25, 26, 27,
 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 37, 38, 39, 40, 41, 42,
 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 56, 57,
 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73,
 74, 75, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88,
 89, 90, 91, 92, 93, 94, 94, 95, 96, 97, 98, 99,100,101,102,103,
241,242,243,244,245,246,247,248,248,  0,  0,  0,  0,  0,  0,  0,
alternative state transition table
The alternative state transition table has been build using iterative minimization of frame sizes and generally performs better than the default. To use it, the coder_type has to be set to 2 and the difference to the default has to be stored in the header. The reference implemenation of FFV1 in FFmpeg uses this table by default at the time of this writing when Range coding is used.
  0, 10, 10, 10, 10, 16, 16, 16, 28, 16, 16, 29, 42, 49, 20, 49,
 59, 25, 26, 26, 27, 31, 33, 33, 33, 34, 34, 37, 67, 38, 39, 39,
 40, 40, 41, 79, 43, 44, 45, 45, 48, 48, 64, 50, 51, 52, 88, 52,
 53, 74, 55, 57, 58, 58, 74, 60,101, 61, 62, 84, 66, 66, 68, 69,
 87, 82, 71, 97, 73, 73, 82, 75,111, 77, 94, 78, 87, 81, 83, 97,
 85, 83, 94, 86, 99, 89, 90, 99,111, 92, 93,134, 95, 98,105, 98,

3.6.2 Huffman coding mode

This coding mode uses golomb rice codes. The VLC code is split into 2 parts, the prefix stores the most significant bits, the suffix stores the k least significant bits or stores the whole number in the ESC case. The end of the bitstream (of the frame) is filled with 0-bits so that the bitstream contains a multiple of 8 bits.
bits value
1 0
01 1
... ...
0000 0000 0001 11
0000 0000 0000 ESC
non ESC the k least significant bits MSB first
ESC the value - 11, in MSB first order, ESC may only be used if the value cannot be coded as non ESC
k bits value
0 1 0
0 001 2
2 1 00 0
2 1 10 2
2 01 01 5
any 000000000000 10000000 139
Run mode
Run mode is entered when the context is 0, and left as soon as a non-0 difference is found, the level is identical to the predicted one, the run and the first different level is coded.
Run length coding
The run value is encoded in 2 parts, the prefix part stores the more significant part of the run as well as adjusting the run_index which determines the number of bits in the less significant part of the run. The 2nd part of the value stores the less significant part of the run as it is. The run_index is reset for each plane and slice to 0.
 0, 0, 0, 0, 1, 1, 1, 1,
 2, 2, 2, 2, 3, 3, 3, 3,
 4, 4, 5, 5, 6, 6, 7, 7,
 8, 9,10,11,12,13,14,15,

if (run_count == 0 && run_mode == 1) {
    if (get_bits1()) {
        run_count = 1 << log2_run[run_index];
        if (x + run_count <= w)
    } else {
        if (log2_run[run_index])
            run_count = get_bits(log2_run[run_index]);
            run_count = 0;
        if (run_index)
        run_mode = 2;
Level coding
Level coding is identical to the normal difference coding with the exception that the 0 value is removed as it cannot occur:
if(diff>0) diff--;
Note, this is different from JPEG-LS, which doesn’t use prediction in run mode and uses a different encoding and context model for the last difference On a small set of test samples the use of prediction slightly improved the compression rate.

4 Bitstream

u(n) unsigned big endian integer using n bits
sg Golomb Rice coded signed scalar symbol coded with the method described in3.6.2↑
br Range coded boolean (1-bit) symbol with the method described in 1↑
ur Range coded unsigned scalar symbol coded with the method described in 2↑
sr Range coded signed scalar symbol coded with the method described in 2↑
The same context which is initialized to 128 is used for all fields in the header.

4.1 Frame

Frame { type
    keyframe br
    if (keyframe) {
    for(i=0; i<slice_count; i++)

4.2 Slice

Slice(i) { type
    if(version>2) {
        slice_x ur
        slice_y ur
        slice_width-1 ur
        slice_height-1 ur
        for(j=0; j<plane_count; j++)
            quant_table_index[i][j] ur
        picture_structure ur
        sar_num ur
        sar_den ur
        if (version > 3)
            reset_contexts br
            slice_coding_mode ur
    if (colorspace_type == 1) {
        for (y=0; y<height; y++) {
            if (alpha_plane)
    } else {
        if (chroma_planes) {
        if (alpha_plane)
    if(i || version>2)
        slice_size u(24)
        error_status u(8)
        crc_parity u(32)
slice_coding_mode indicates the slice coding mode.
value slice coding mode
0 normal Range Coding or VLC
1 raw PCM
Other reserved for future use
slice_size indicates the size of the slice in bytes.
Note: this allows finding the start of slices before previous slices have been fully decoded. And allows this way parallel decoding as well as error resilience.
error_status specifies the error status.
value error status
0 no error
1 slice contains a correctable error
2 slice contains a uncorrectable error
Other reserved for future use
plane_count indicates the count of planes and the associated plane types.
value plane types
0 forbidden
1 if version <4: forbidden; else gray
2 if version <4: forbidden; else gray+alpha
3 luma+chroma
4 luma+chroma+alpha
Other reserved for future use

4.3 Header

4.3.1 Version 0 and 1

FrameHeader01 { type
    version ur
    coder_type ur
        for(i=1; i<256; i++)
            state_transition_delta[i] sr
    colorspace_type ur
        bits_per_raw_sample ur
    chroma_planes br
    log2(h_chroma_subsample) ur
    log2(v_chroma_subsample) ur
    alpha_plane br

4.3.2 Version 3

Version 2 and later files use a global header and a per frame header.
GlobalHeader { type
    version ur
    micro_version ur
    coder_type ur
        for(i=1; i<256; i++)
            state_transition_delta[i] sr
    colorspace_type ur
    bits_per_raw_sample ur
    chroma_planes br
    log2(h_chroma_subsample) ur
    log2(v_chroma_subsample) ur
    alpha_plane br
    num_h_slices-1 ur
    num_v_slices-1 ur
    quant_table_count ur
    for(i=0; i<quant_table_count; i++)
    for(i=0; i<quant_table_count; i++) {
        states_coded br
            for(j=0; j<context_count[i];j++)
                for(k=0; k<CONTEXT_SIZE;k++)
                    initial_state_delta[i][j][k] sr
    ec ur
    intra ur
    crc_parity u(32)
version specifies the version of the bitstream.
Each version is incompatible with others versions: decoders SHOULD reject a file due to unknown version.
value version
0 FFV1 version 0
1 FFV1 version 1
2 reserved*
3 FFV1 version 3
Other reserved for future use

* Version 2 was never enabled in the encoder thus version 2 files SHOULD NOT exist, and this document does not describe them to keep the text simpler.
micro_version specifies the micro-version of the bitstream.
After a version is considered stable (a micro-version value is assigned to be the first stable variant of a specific version), each new micro-version after this first stable variant is compatible with the previous micro-version: decoders SHOULD NOT reject a file due to an unknown micro-version equal or above the micro-version considered as stable.

Meaning of micro_version for version 3:
value micro_version
0...3 reserved*
4 first stable variant
Other reserved for future use

* were development versions which may be incompatible with the stable variants.

Meaning of micro_version for version 4 (note: at the time of writting of this specification, version 4 is not considered stable so the first stable version value is to be annonced in the future):
value micro_version
0...TBA reserved*
TBA first stable variant
Other reserved for future use

* were development versions which may be incompatible with the stable variants.
coder_type specifies the coder used
value coder used
0 Golomb Rice
1 Range Coder with default state transition table
2 Range Coder with custom state transition table
Other reserved for future use
state_transition_delta specifiies the range coder custom state transition table.
If state_transition_delta is not present in the bitstream, all range coder custom state transition table elements are assumed to be 0.
colorspace_type specifies the color space.
value color space used
0 YCbCr
1 JPEG 2000 RCT
Other reserved for future use
chroma_planes indicates if chroma (color) planes are present.
value color space used
0 chroma planes are not present
1 chroma planes are present
bits_per_raw_sample indicates the number of bits for each luma and chroma sample.
Inferred to be 8 if not present.
value bits for each luma and chroma sample
0 reserved*
Other the actual bits for each luma and chroma sample

* Encoders MUST not store bits_per_raw_sample = 0
Decoders SHOULD accept and interpret bits_per_raw_sample = 0 as 8.
h_chroma_subsample indicates the subsample factor between luma and chroma width (chroma_width = 2 − log2_h_chroma_subsampleluma_width)
v_chroma_subsample indicates the subsample factor between luma and chroma height (chroma_height = 2 − log2_v_chroma_subsampleluma_height)
alpha_plane indicates if a transparency plane is present.
value color space used
0 transparency plane is not present
1 transparency plane is present
num_h_slices indicates the number of horizontal elements of the slice raster.
num_v_slices indicates the number of vertical elements of the slice raster.
quant_table_count indicates the number of quantization table sets.
states_coded indicates if the respective quantization table set has the initial states coded.
value initial states
0 initial states are not present and are assumed to be all 128
1 initial states are present
initial_state_delta[i][j][k] indicates the initial range coder state, it is encoded using k as context index and
pred= j ? initial_states[i][j-1][k] : 128
initial_state[i][j][k]= (pred+initial_state_delta[i][j][k])&255
slice_count indicates the number of slices in the current frame, slice_count is 1 if it is not explicitly coded.
slice_x indicates the x position on the slice raster formed by num_h_slices.
slice_y indicates the y position on the slice raster formed by num_v_slices.
slice_width indicates the width on the slice raster.
slice_height indicates the height on the slice raster.
quant_table_index indicates the index to select the quantization table set and the initial states for the slice.
picture_structure specifies the picture structure.
value picure structure used
0 unknown
1 top field first
2 bottom field first
3 progressive
Other reserved for future use
sar_num specifies the sample aspect ratio numerator.
MUST be 0 if sample aspect ratio is unknown.
sar_den specifies the sample aspect ratio numerator.
MUST be 0 if sample aspect ratio is unknown.
ec indicates the error detection/correction type.
value error detection/correction type
0 32bit CRC on the global header
1 32bit CRC per slice and the global header
Other reserved for future use
intra indicates the relationship between frames.
value relationship
0 frames are independent or dependent (key and non key frames)
1 frames are independent (key frames only)
Other reserved for future use
crc_parity 32bit that are choosen so that the global header as a whole or slice as a whole has a crc remainder of 0. This is equivalent to storing the crc remainder in the 32bit parity. The CRC generator polynom used is the standard IEEE CRC polynom (0x104C11DB7) with initial value 0. The compressed bitstream is padded so that the 32bit crc end in the last 4 bytes.

4.4 Quantization Tables

The quantization tables are stored by storing the number of equal entries -1 of the first half of the table using the method described in 2↑. The second half doesn’t need to be stored as it is identical to the first with flipped sign
Table: 0 0 1 1 1 1 2 2-2-2-2-1-1-1-1 0
Stored values: 1, 3, 1
QuantizationTable (i) { type
    for(j=0; j<MAX_CONTEXT_INPUTS; j++) {
        QuantizationTablePerContext(i, j, scale)
QuantizationTablePerContext (i, j, scale) { type
    for(k= 0; k<128;) {
        len-1 sr
        for(a=0; a<len; a++) {
    for (k=1; k<128; k++) {
quant_tables indicates the quantification table values.
context_count indicates the count of contexts.

4.5 Restrictions

In version 2 and later the maximum slice size in pixels is (widthheight)/(4), this is to ensure that fast multithreaded decoding is possible.

5 Changelog

See https://github.com/FFmpeg/FFV1/commits/master

6 ToDo

7 Bibliography


[1] JPEG-LS FCD 14495 http://www.jpeg.org/public/fcd14495p.pdf

[2] H.264 Draft http://bs.hhi.de/~wiegand/JVT-G050.pdf

[3] HuffYuv http://cultact-server.novi.dk/kpo/huffyuv/huffyuv.html

[4] FFmpeg http://ffmpeg.org

[5] JPEG2000 http://www.jpeg.org/jpeg2000/

[6] "Range encoding: an algorithm for removing redundancy from a digitised message. Presented by G. Nigel N. Martin at the Video & Data Recording Conference, IBM UK Scientific Center held in Southampton July 24-27 1979."

8 Copyright

Copyright 2003-2013 Michael Niedermayer <michaelni@gmx.at>
This text can be used under the GNU Free Documentation License or GNU General Public License. See http://www.gnu.org/licenses/fdl.txt.