Quality assurance · Production

Netflix improves streaming video quality with neural network-based deep downscaler

The problem

Netflix's video encoding pipeline relied on conventional resampling filters like Lanczos for video downscaling, limiting achievable quality across device resolutions and network conditions.

Workflow diagram · grounded in source
1
Source video enters pipeline
trigger
“Video downscaling is the most pertinent example herein, which tailors our encoding to screen resolutions of different devices and optimizes picture quality under varying network conditions”
2
NN preprocessing block
ai_action
“The preprocessing block aims to prefilter the video signal prior to the subsequent resizing operation”
3
NN resizing block
ai_action
“The resizing block yields the lower-resolution video signal that serves as input to an encoder”
4
Conventional codec encoding
integration
“Video encoding using a conventional video codec, like AV1”
5
Streaming to member devices
output
“Millions of devices that support Netflix streaming automatically benefit from this solution”
Reported outcome

The deep downscaler achieved a ~5.4% BD rate gain over Lanczos for VP9 encoding and was preferred by ~77% of human test subjects, with A/B testing confirming QoE improvements without adverse streaming impact.

Reported metrics
VMAF BD rate gain (VP9 over Lanczos)~5.4%
VMAF-NEG BD rate gain~4.4%
Human subject preference rate~77%
streaming QoE impactQoE improvements without any adverse streaming impact
Reported stack
FFmpegCosmosTitus
Source
https://netflixtechblog.com/for-your-eyes-only-improving-netflix-video-quality-with-neural-networks-5b8d032da09c
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The deep downscaler achieved a ~5.4% BD rate gain over Lanczos for VP9 encoding and was preferred by ~77% of human test subjects, with A/B testing confirming QoE improvements without adverse streaming impact.

What tools did this team use?

FFmpeg, Cosmos, Titus.

What results were reported?

VMAF BD rate gain (VP9 over Lanczos): ~5.4%; VMAF-NEG BD rate gain: ~4.4%; Human subject preference rate: ~77%; streaming QoE impact: QoE improvements without any adverse streaming impact (source-reported, not independently verified).

How is this quality assurance AI workflow structured?

Source video enters pipeline → NN preprocessing block → NN resizing block → Conventional codec encoding → Streaming to member devices.

Related quality assurance cases