Note that for this, you should always use the latest ffmpeg version, and preferably compile it yourself. This gives you access to the most recent libx265 and libfdk-aac for audio encoding.
Also, the data rate savings will be quite drastic if you're going from a ~10 MBit/s DVD to around 1–2 MBit/s for H.264 video and 0.5–1 MBit/s for H.265 video. Changing the quality in the below steps may influence the bitrates, but still the data reduction should be significant.
H.264
For the quality / rate control, you want to use CRF mode in libx264 rather than a constant bitrate. Using CRF ensures that an average quality is preserved, independent of the original video resolution or its complexity. Constant bitrate is only really useful if you're constrained by the transmission medium (e.g. hard drive speed, Internet throughput).
Choosing the CRF value is the tricky part. It requires you to look at the output. The default for libx264 (23) offers a quite good tradeoff between size and quality. But given that your original source is already compressed (and not with a very good quality compared to Blu-rays), you may want to change the CRF to be a little lower, such as 20. This will increase needed bitrate by about a third.
Choose the preset according to how long you want to wait. slow
seems like a good value here.
ffmpeg -i input \
-c:v libx264 -crf 20 -pix_fmt yuv420p \
-x264-params keyint=240:min-keyint=20 \
-preset:v slow -profile:v baseline -level 3.0 \
-c:a libfdk_aac -vbr 4 \
output.mp4
The built-in ffmpeg AAC encoder can be used if libfdk-aac is not available. Use -c:a aac -strict experimental -b:a 128k
instead of -c:a libfdk_aac -vbr 4
.
H.265
Research suggests that using HEVC will lead to up to 74% bitrate saving compared to H.264. This is based on subjective viewing data of Ultra-HD sequences. Of course, it depends on the temporal complexity of the source content, and the amount of data saved will not be as high for hard to code sequences. Either way you can safely say that 50% data reduction is absolutely possible.
The default CRF for libx265 is 28. Using the same source content, it results in about half the bitrate compared to libx264 at CRF 23. This is irrespective of the actual bitrate, i.e., if the H.264 version takes 1.5 MBit/s, then H.265 will use around 750 kBit/s, but it's 750 kBit/s vs. 350 kBit/s for another sequence. I ran it on a couple of sequences at DVD-PAL resolution and was not able to tell the difference in terms of quality.
ffmpeg -i input \
-c:v libx265 -pix_fmt yuv420p \
-x265-params crf=28:keyint=240:min-keyint=20 \
-preset:v slow \
-c:a libfdk_aac -vbr 4 \
output.mp4
For more information, here are the relevant resources:
Thank you for a fine answer. What does keyint mean practically by the way? – Ivan – 2014-10-17T14:16:44.100
1The
keyint
in x264/x265 is the interval between IDR-frames, i.e. the interval between keyframes at which the decoder can refresh. In between, there can be non-keyframe I-frames, e.g. when a scene cut occurs. It's equivalent to the-g
parameter if I'm not mistaken. – slhck – 2014-10-17T14:21:31.030BTW, @slhck, a thing that has surprised me in your answer - the attention you give to the choice of an AAC encoding library. I used to think they are all almost the same and give little or no difference, that the things are simple in the audio part (just choose the bitrate and go and that all the major lossy codecs like MP3, AAC and Vorbis sound almost or exactly the same at 128 kbps and above). Do you mean there actually is notable difference between libfdk-aac and ordinary aac? – Ivan – 2014-10-17T16:21:13.793
+1 on mentioning recent builds since old ones may not set the correct number of
-refs
for libx264 as defined by the desired-level
(old bug 3307, and one other one IIRC). – llogan – 2014-10-17T16:37:06.050Well @Ivan, it depends on how much of an audiophile you are or how much attention you pay to the audio quality. In fact, like you mention, at some bitrate threshold there will be no difference between the encoders, but I'm trying to encourage people to use libfdk whenever possible since the libvo-aacenc that you get with most ffmpeg builds has really bad quality (Japanese only, sorry). The built-in one is nice but doesn't do VBR.
– slhck – 2014-10-17T16:44:12.623I am using the latest Zeranoe FFmpeg Win32 builds from http://ffmpeg.zeranoe.com/builds/ and the only way I have managed to get AAC out of it was
– Ivan – 2014-10-17T18:33:32.453-c:a libvo_aacenc
. I would try building it on my own if I were running GNU/Linux but am afraid to imagine building software from sources (other than a VisualStudio project) on Windows...1@Ivan The Zeranoe builds should definitely allow you to do
-c:a aac -strict experimental
as indicated in my answer. And I agree, I wouldn't try to build it on Windows. – slhck – 2014-10-17T18:34:56.1331
@Ivan (1st comment): See ffmpeg-wiki: "Based on quality produced from high to low:
– Golar Ramblar – 2016-07-12T11:03:52.473libopus > libvorbis >= libfdk_aac > aac > libmp3lame >= libfaac >= eac3/ac3 > libtwolame > vorbis > mp2 > wmav2/wmav1
For AAC only: (Because it is a little bit confusing, with 3 encoders available):libfdk_aac > aac > libfaac
The >= sign means greater or the same quality."