Some Quick Takes on 8K “Per-Title” Encoding
Article co-authored by Ben Schwarz. When encoding guru Jan Ozer gave a seminar on “Per-Title Encoding”, we took notice and contacted him regarding the state of such practices with 8K. He noted that he had no data on 8K titles, so two short pieces of content were provided for him to play with. He has posted his first thoughts on his website, and the results also inspired some conclusions of our own.
As he details in his article, the first piece of content was 8K @ 25 fps in an SDR and rec. 709 format. Selected clips included a stationary talking head scene along with a dance sequence, which showed a lot of motion blur due to the capture frame rate. The second clip was 8K at 60 fps in an HDR and BT.2020 format. A dynamic action sequence was chosen for encoding.
All test encodes used HEVC with a CRF 25 quality level, adjusting the encoding bit rate to achieve the desired quality level. Both PSNR and SSIM were measured to assess quality.
- Per-title encoding will be critical to achieving manageable bit rates for 8K content. “per-title encoding” is a Netflix term that refers to creating encoding parameters on a scene-by-scene basis. Before introducing this approach, encoding typically used the same parameters over an entire piece of content, which can waste a lot of bits. Other techniques can encode on a frame basis. “Content-Aware Encoding” is perhaps a more appropriate term that has been adopted by industry bodies such as the Ultra HD Forum that even goes as far as saying that CAE is a requirement for the OTT delivery of very high resolutions. Operators that adopted CAE to cope with the spike in binge-watching during the first lockdown last spring saw reductions from 25% to 40% in streaming bandwidth.
- 8K SDR low motion images at 25 FPS can be encoded at very reasonable data rates – 2.46 Mbps with an SSIM score of 0.993. This level means the encoded content should be visually unchanged from the original. Neither Ozer nor the 8K Association expected such a low data rate would be possible.
- 8K low frame rate images with motion blur can also be encoded at very reasonable data rates. The dance sequence came in at 7.8 Mbps with a 0.996 SSIM score.
- The 8K/60 HDR soccer sequence used 43.4 Mbps, reaching a 0.998 SSIM score, consistent with what Harmonic has seen on its encoding trials with this content. This rate is for single-pass encoding to support live 8K. Multiple passes should be able to lower this rate. And, as AI/ML techniques improve, commercially acceptable data rates for 8K might not be that far off.
The tests also suggested some additional questions, which we may explore in future posts.
- Would there be a big difference between the SDR and HDR versions of a scene in terms of bit-rate and metrics? Apple has suggested the delta is 20%, but others think it lower. However, such data is not based on 8K content and at 30 fps or lower. Would it be different for 8K/30? 8K/60?
- PSNR scores do not match the SSIM scores. Is PSNR relevant anymore? Have other qualitative metrics like SSIM and VMAF been validated for 8K HDR content? We don’t think so.