I’ve been in the streaming media industry since 2008 and have seen a lot of misinformation regarding both FFmpeg and libx264. In this post I hope to help shed some light on what does and does not work.
Streaming media, at it’s core, requires three basic things.
1) Constant frame rate.
2) An even keyframe distance which is also known as a Group of Pictures or GOP.
3) A bitrate based encode.
Things that are nice to have are.
4) Finding a better bitrate for your content.
5) Hitting your target bitrate.
6) Audio encoding without A/V drift.
7) Proper encoding for your target audiences.
I have a basic rule when encoding content and that is to never trust the input. Are you sure that the frame rate is constant? Were you told that the content is progressive and not interlaced? Were you given information about keyframe distance or what color space the video is in? Can you trust that any of that is accurate? I can’t and you shouldn’t.
Constant frame rate is important because players like to have the PTS/DST timestamps they are decoding generated like clockwork. If they are not in the correct order you can have playback problems like content jumping forwards, backwards, and even possible problems with basic playback. To achieve proper playback with FFmpeg you need to use two options.
-r is used to specify the output frame rate. This must be the same as the input frame rate to eliminate judder. This is used in conjunction with the
-vsync parameter using the
1 option which will retime the PTS/DTS timestamps accordingly. Depending upon the content you get do not be surprised if frames are duplicated and/or dropped during encoding. If that happens then if possible contact the content creator and ask them to fix their source content. It is not uncommon for FFmpeg to duplicate the first frame.
Ensuring that your keyframe distance is always the same you can use the
-g parameter. I go a bit beyond what is required for regular desktop playback and use the
no-scenecut option in conjunction with the
-g parameter. x264 will, by default, create a keyframe when it detects a scene change. It will also set the default maximum GOP value to 250 and the minimum GOP value to 25. Using the
no-scenecut option will turn off scene detection for that codec. Setting
--scenecut -1 is not a valid option or if it is I have found it nowhere in either x264’s or FFmpeg’s documentation.
Note that NTSC content is stupid. 23.976fps is inaccurate and should always be written as 24000/1001. 29.970fps should always be written as 30000/1001. 59.94fps should always be written as 60000/1001. The examples below are inaccurate on purpose for ease of reading.
If you inspect an output file with MediaInfo and did not use the
no-scenecut option you will see
scenecut=40. When done properly that will be zero
scenecut=0. If this option is not used then keyframes will be misaligned for ABR content and segment sizes will be unpredictable.
You can also use the FFmpeg
-sc_threshold 0 parameter to disable scene detection and is video codec neutral. This is equivalent to the
no-scenecut option provided by libx264.
I have seen people attempt to create VOD content and perform live streaming using Constant Rate Factor which is also known as
CRF. If you do not specify a bitrate for x264 then it will default to
CRF 23. If you do not specify a preset it will default to medium. If you do not specify a profile it will default to high.
I like to make sure that my content uses all of the bells and whistles for delivering bitrate based content including
-maxrate, and even
-minrate. Note that
-minrate has no effect with x264. It is in my script in case I decide to use a different codec and it supports a minimum bitrate. [Note to self: update script to use
-sc_threshold 0 ]
If you want to take the guesswork out of finding a better bitrate then it is best to analyze the file to find one. I now use
CRF 23 to find a better global bitrate for whatever I am encoding and make sure to use the same encoding settings as my output file with the exception that I use the
veryfast preset and the
baseline profile. This is vital to finding a better bitrate.
Below is a sample of my script that I use for two pass encoding. Note that almost everything in the script is a variable. Those values are inserted after the media is analyzed and a better bitrate is detected as described in the bitrate detection section above. I also perform audio conversion separate from video conversion as encoding audio at the same time can slow down this process.
ffmpeg -i $inputfile $scan -pix_fmt $colorspace -vf "crop=$w1:$h1:$x1:$y1,scale=$fixedwidth:$fixedheight" -vsync 1 -sn -map $vtrack -r $fps -threads 0 -vcodec libx264 -b:v:$vtrack $averagevideobitrate -bufsize $buffer -maxrate $maximumvideobitrate -minrate $minimumvideobitrate -an -pass 1 -preset $newpreset -profile:v $defaultprofile -g $gop $tune -x264opts no-scenecut -map_metadata -1 -f mp4 -y $outputfile-video.mp4
ffmpeg -i $inputfile $scan -pix_fmt $colorspace -vf "crop=$w1:$h1:$x1:$y1,scale=$fixedwidth:$fixedheight" -vsync 1 -sn -map $vtrack -r $fps -threads 0 -vcodec libx264 -b:v:$vtrack $averagevideobitrate -bufsize $buffer -maxrate $maximumvideobitrate -minrate $minimumvideobitrate -an -pass 2 -preset $newpreset -profile:v $defaultprofile -g $gop $tune -x264opts no-scenecut -map_metadata -1 -f mp4 -y $outputfile-video.mp4
The same values are used for the second pass to ensure that target bitrate is hit. If you do not use the same parameters in both passes then you will always miss your target bitrate.
This is an example of bad two pass encoding where different values are used, between the two passes neither frame rate nor GOP are defined, and your PTS/DTS timestamps will be the same as the input. You will never hit your target bitrate using this method.
ffmpeg -y -i 1080p-input.mp4 -c:v libx264 -b:v 5000k -pass 1 -f mp4 NUL && \
ffmpeg -i 1080p-input.mp4 -c:v libx264 -b:v 5000k -maxrate 5000k -bufsize 5000k -pass 2 1080p-output.mp4
Never reuse your first pass analysis when creating Adaptive Bitrate (ABR) content. Ever.
ffmpeg -y -i 1080p-input.mp4 -c:v libx264 -preset medium -g 60 -keyint_min 60 -sc_threshold 0 -bf 3 -b_strategy 2 -b:v 3000k -c:a aac -b:a 64k -ac 1 -ar 44100 -pass 1 -f mp4 NUL && \
ffmpeg -i 1080p-input.mp4 -c:v libx264 -preset medium -g 60 -keyint_min 60 -sc_threshold 0 -bf 3 -b_strategy 2 -b:v 3000k -maxrate 3300k -bufsize 3000k -c:a aac -b:a 64k -ac 1 -ar 44100 -pass 2 1080p-output.mp4
ffmpeg -i 1080p-input.mp4 -c:v libx264 -s 1280x720 -preset medium -g 60 -keyint_min 60 -sc_threshold 0 -bf 3 -b_strategy 2 -b:v 1500k -maxrate 1650k -bufsize 1500k -c:a aac -b:a 64k -ac 1 -ar 44100 -pass 2 720p_output.mp4
ffmpeg -i 1080p-input.mp4 -c:v libx264 -s 640x360 -preset medium -g 60 -keyint_min 60 -sc_threshold 0 -bf 3 -b_strategy 2 -b:v 1000k -maxrate 1100k -bufsize 1000k -c:a aac -b:a 64k -ac 1 -ar 44100 -pass 2 360p-output.mp4
My personal experience using
-b_strategy 2 did not work out so well and actually lowered the quality of my content. Your mileage may vary. Using
-bf 3 will force three B-frames to be used. This is the default in the medium preset. In addition the medium preset uses three reference frames for content. This is easy for today’s players to decode.
This is two pass encoding done right while also converting audio to stereo AAC. I include the
pix_fmt yuv420p color space because if you convert a piece of content that has, say, an incompatible color space (See also desktop Windows Media content) or is using the color range of computer RGB (0-255) and not broadcast RGB (16-235), then your H.264 video may not play back as expected.
ffmpeg -i inputfile.mp4 -pix_fmt yuv420p -vsync 1 -vcodec libx264 -r 23.976 -threads 0 -b:v: 1024k -bufsize 1216k -maxrate 1280k -preset medium -profile:v high -tune film -g 48 -x264opts no-scenecut -pass 1 -acodec aac -b:a 192k -ac 2 -ar 48000 -af "aresample=async=1:min_hard_comp=0.100000:first_pts=0" -f mp4 -y outputfile.mp4
ffmpeg -i inputfile.mp4 -pix_fmt yuv420p -vsync 1 -vcodec libx264 -r 23.976 -threads 0 -b:v: 1024k -bufsize 1216k -maxrate 1280k -preset medium -profile:v high -tune film -g 48 -x264opts no-scenecut -pass 2 -acodec aac -b:a 192k -ac 2 -ar 48000 -af "aresample=async=1:min_hard_comp=0.100000:first_pts=0" -f mp4 -y outputfile.mp4
Note that I add the audio bitrate to the video bitrate to calculate the
bufsize value. I also multiply the target bitrate by 1.25 for the
maxrate value. Why? This provides the encoder the liberty to allocate less data to low motion scenes and more data to higher action scenes. If you were to use a 10x value for your
maxrate value the network signature would look a lot like CRF but boy will your content look great. I do not recommend this.
This brings up a few questions regarding quality, compression, and two pass encoding.
1) Is there a visible difference in output between
CRF 23 when using the
veryfast preset and the
baseline profile versus the
medium preset and the
You would think that as
CRF 23 is being used that both output videos would be the same quality. Unfortunately this does not appear to be the case and Moscow State Universities’ Video Quality Metric Tool confirms this when analyzing the two output files via SSIM. I would use the Netflix VMAF tool but it is segfaulting as of the time I am writing this if it is both included with FFmpeg and is analyzing content via SSIM.
On a side note B-frames are killing me with regards to quality. While the output size of the high profile is smaller so too is it’s bit per pixel density. I interpret this as both good and bad. Compression, at least in this case, costs quality but it does make the file smaller.
Veryfast preset with the baseline profile using CRF 23:
Size == 888 MiB "Bitrate" == 1023 kb/s BPP == 0.104
Medium preset with the high profile using CRF 23:
Size == 833 MiB "Bitrate" == 958 kb/s BPP == 0.098
This lowered SSIM quality to an average of 0.97495 between the two files.
2) Does using the 1080p two pass mbtree file and it’s two pass log file for encoding other pieces of content degrade quality?
Yes. Now why did I run that test? Because a lot of people reuse the first pass files for their other outputs in their ABR stack as shown earlier in this article. I have never agreed with that so I did a direct compare with a properly encoded two pass file and then used the 1080p mbtree and log file to output a 480p file.
To bring this home let us take a look at what happens when you use the proper two pass log files versus what happens when you use the wrong ones. In this instance the source content was 1080p and was used to generate the
ffmpeg2pass-0.log.mbtree file and the
ffmpeg2pass-0.log file. A second two pass encode was used to create a 480p output from the same 1080p source just like you would do when creating ABR content.
The two pass log file size for the 1080p mbtree file weighed in at 1.40GB while the 480p mbtree file weighed in at 287MB.
This is the second pass of the 480p output using the proper mbtree and log files.
frame=174434 fps=104 q=-1.0 Lsize= 907956kB time=02:01:15.23 bitrate=1022.4kbits/s dup=0 drop=1 speed=4.32x video:905937kB audio:0kB subtitle:0kB other streams:0kB global headers:0kB muxing overhead: 0.222847% [libx264 @ 000000000305b2e0] frame I:3635 Avg QP:18.51 size: 31828 [libx264 @ 000000000305b2e0] frame P:47427 Avg QP:22.94 size: 9521 [libx264 @ 000000000305b2e0] frame B:123372 Avg QP:24.57 size: 2922 [libx264 @ 000000000305b2e0] consecutive B-frames: 4.3% 1.9% 7.1% 86.7% [libx264 @ 000000000305b2e0] mb I I16..4: 23.8% 52.0% 24.2% [libx264 @ 000000000305b2e0] mb P I16..4: 2.1% 8.6% 2.7% P16..4: 32.3% 12.3% 8.0% 0.0% 0.0% skip:34.1% [libx264 @ 000000000305b2e0] mb B I16..4: 0.1% 0.7% 0.2% B16..8: 35.7% 4.2% 0.9% direct: 2.0% skip:56.2% L0:44.0% L1:49.2% BI: 6.9% [libx264 @ 000000000305b2e0] 8x8 transform intra:60.9% inter:72.3% [libx264 @ 000000000305b2e0] coded y,uvDC,uvAC intra: 65.0% 60.8% 30.4% inter: 16.0% 14.5% 0.5% [libx264 @ 000000000305b2e0] i16 v,h,dc,p: 44% 30% 10% 16% [libx264 @ 000000000305b2e0] i8 v,h,dc,ddl,ddr,vr,hd,vl,hu: 17% 20% 18% 6% 7% 8% 8% 7% 8% [libx264 @ 000000000305b2e0] i4 v,h,dc,ddl,ddr,vr,hd,vl,hu: 21% 25% 15% 5% 7% 7% 7% 5% 6% [libx264 @ 000000000305b2e0] i8c dc,h,v,p: 58% 21% 16% 5% [libx264 @ 000000000305b2e0] Weighted P-Frames: Y:13.2% UV:3.7% [libx264 @ 000000000305b2e0] ref P L0: 52.5% 16.0% 20.8% 9.4% 1.4% [libx264 @ 000000000305b2e0] ref B L0: 84.5% 12.2% 3.3% [libx264 @ 000000000305b2e0] ref B L1: 94.6% 5.4% [libx264 @ 000000000305b2e0] kb/s:1020.08
This is the second pass of the 480p output using incorrect 1080p mbtree and log files.
frame=174434 fps=106 q=-1.0 Lsize= 907905kB time=02:01:15.23 bitrate=1022.3kbits/s dup=0 drop=1 speed=4.41x video:905883kB audio:0kB subtitle:0kB other streams:0kB global headers:0kB muxing overhead: 0.223274% [libx264 @ 00000000026b00a0] frame I:3635 Avg QP:20.20 size: 26166 [libx264 @ 00000000026b00a0] frame P:46748 Avg QP:22.62 size: 9715 [libx264 @ 00000000026b00a0] frame B:124051 Avg QP:24.55 size: 3050 [libx264 @ 00000000026b00a0] consecutive B-frames: 4.0% 1.4% 6.8% 87.8% [libx264 @ 00000000026b00a0] mb I I16..4: 22.9% 52.5% 24.5% [libx264 @ 00000000026b00a0] mb P I16..4: 2.3% 9.0% 3.2% P16..4: 31.0% 11.7% 7.4% 0.0% 0.0% skip:35.5% [libx264 @ 00000000026b00a0] mb B I16..4: 0.1% 0.7% 0.2% B16..8: 33.7% 4.3% 0.9% direct: 2.4% skip:57.7% L0:41.7% L1:50.1% BI: 8.2% [libx264 @ 00000000026b00a0] 8x8 transform intra:60.0% inter:69.7% [libx264 @ 00000000026b00a0] coded y,uvDC,uvAC intra: 62.2% 59.8% 30.0% inter: 15.6% 14.1% 0.8% [libx264 @ 00000000026b00a0] i16 v,h,dc,p: 41% 28% 11% 20% [libx264 @ 00000000026b00a0] i8 v,h,dc,ddl,ddr,vr,hd,vl,hu: 18% 21% 17% 6% 7% 8% 8% 7% 8% [libx264 @ 00000000026b00a0] i4 v,h,dc,ddl,ddr,vr,hd,vl,hu: 20% 28% 16% 5% 7% 7% 7% 5% 5% [libx264 @ 00000000026b00a0] i8c dc,h,v,p: 59% 21% 16% 5% [libx264 @ 00000000026b00a0] Weighted P-Frames: Y:8.8% UV:2.1% [libx264 @ 00000000026b00a0] ref P L0: 52.2% 16.4% 21.2% 9.4% 0.9% [libx264 @ 00000000026b00a0] ref B L0: 84.3% 12.3% 3.3% [libx264 @ 00000000026b00a0] ref B L1: 94.6% 5.4% [libx264 @ 00000000026b00a0] kb/s:1020.02
Take note that a lot of data was pulled out of the I frames and allocated to P and B frames when the 1080p
ffmpeg2pass-0.log.mbtree file and the
ffmpeg2pass-0.log file were used instead of the ones that were generated for the output 480p content. This lowered SSIM quality to an average of 0.97553 when using the wrong two pass files. Viewing some of the frame differences in MSU VQMT made my eyes hurt.
Did you notice that the bit rate of the 480p file that was encoded using the
veryfast preset, the
baseline profile, and
CRF 23 is very close (within the bounds of CAVLC and CAVAC entropy encoding) to the bitrate of the two pass encode that used the medium preset and the high profile? Two pass encoding puts the bits back.
In the audio portion of the two lines above you will see a few filters.
-af "aresample=async=1:min_hard_comp=0.100000:first_pts=0" helps to keep your audio lined up with the beginning of your video. It is common for a container to have the beginning of the video and the beginning of the audio start at different points. By using this your container should have little to no audio drift or offset as it will pad the audio with silence or trim audio with negative PTS timestamps if the audio does not actually start at the beginning of the video.
During my nine years in the streaming media industry I have seen companies like Research in Motion state emphatically that their Blackberry phones do not support the
high profile. Just because a manufacturer states that specific profiles are not supported does not mean that they won’t work. RealNetwork’s Helix Producer would create, by default, content using the
high profile. I had never had a problem delivering
high profile content to those phones via RTSP.
Generally speaking I limit reference frames to three for compatibility purposes. If you decide to go above that make sure that you research which device or devices support larger reference frame distances. Note that using the
animation tuning option for x264 will double your reference frames unless it is set to one.
Let’s finish off with one final example using single pass encoding..
ffmpeg -i inputfile.mp4 -pix_fmt yuv420p -deinterlace -vf "scale=640:360" -vsync 1 -vcodec libx264 -r 29.970 -threads 0 -b:v: 1024k -bufsize 1216k -maxrate 1280k -preset medium -profile:v main -tune film -g 60 -x264opts no-scenecut -acodec aac -b:a 192k -ac 2 -ar 44100 -af "aresample=async=1:min_hard_comp=0.100000:first_pts=0" -f mp4 -y outputfile.mp4
I hope that this article helps to debunk misinformation that is rampant on the Internet regarding the usage of FFmpeg and what x264 options are valid to create compliant streaming media VOD content.
 Force the output frame rate using the
 Vsync parameter for FFmpeg.
-g option is described below.
 The articles below have several recommendations on bitrate.
 I have an article on how to calculate a better bitrate.
The article above is featured in an article by Jan Ozer. It is also referenced in his book Video Encoding by the Numbers in Chapter 7: Choosing Data Rate.
 Please note that both the first pass and the second pass are identical sans the portion that identifies which pass is being used.
 Detailed information about the macroblock tree file can be found in the following forum by the person who designed it.
 x264 tuning values.
-i is for designating the input.
-deinterlace should only be used if your content is interlaced and is announced as either
MBAFF. It is recommended to deliver only progressive content to web based players.
-vf "scale=640:360" is a video filter that will scale the output video to a different resolution.
-vcodec libx264 specifies the x264 video codec. You can substitute -c:v for -vcodec if you wish.
-b:v: 1024k specifies a video bitrate of 1024kbps.
-bufsize 1216k specifies the buffer. This is a best practice for RTSP delivery and streaming media in general.
-maxrate 1280k specifies the maximum bitrate allowed.
-preset veryfast is one of several presets available for H.264 video. Those include ultrafast, superfast, veryfast, faster, fast, medium, slow, slower, veryslow, placebo. It is disrecommended to use anything higher than the medium preset for streaming media.
-profile:v baseline is one of several profiles available for H.264 video. Those include
high444. Hardware devices, specifically older mobile phones, rarely state support for any of the
high profile options even though they may work. You should include the
:v portion at the end of the profile to specify that the profile is for video as some audio codecs also have audio profiles.
Note that x264 has eradicated the
Additional detail on the inner workings of presets, please reference the following page.
-tune film is one of several tuning options available for H.264 video. Those include
Animation should not be used with streaming media as it will double the number of reference frames defined in in the preset option.
More on preset, profile and tuning can be found here.
The libx264 option
ratetol=0.01 will force a very strict constant bitrate, so much so that libx264 will complain and adjust accordingly. This is optional and not shown above as constant bitrate content is dead to me.
-f mp4 defines that the format will be an MP4 container.
-y outputfile.mp4 will state that if outputfile.mp4 exists that it will be overwritten. This is required if you perform two pass encoding and do not redirect the first pass output to a null device.
-acodec aac invokes the use of the internal AAC codec. You can substitute
-acodec if you wish. You no longer need to use the
-strict experimental option with this codec.
-b:a 192k states the the total bitrate of the audio should be 192Kbps. Apple recommends a minimum bitrate of 64Kbps per channel.
-ac 2 forces the audio to be stereo. This is a best practice for streaming media so that you can reach the most players, however you can use additional channels if one or more of your target devices support it.
-ar 44100 forces the frequency to be 44.1k which is compatible with Flash players. The player may downsample audio to 44.1k, 22.05k, or 11.025k. Do not use different audio frequencies with ABR content.
To help translate options between FFmpeg and libx264 please reference the following site.