Netflix patents method to train AI for better video scaling
Netflix describes a technique for training two artificial intelligence models to work together on video quality. One model shrinks an image to a lower resolution, while the other expands it back to the original size. The system adjusts both models simultaneously to minimize the difference between the final image and the original source.
In the filing’s own words
In various embodiments a training application trains convolutional neural networks (CNNs) to reduce reconstruction errors. The training application executes a first CNN on a source image having a first resolution to generate a downscaled image having a second resolution. The training application executes a second CNN on the downscaled image to generate a reconstructed image having the first resolution. The training application computes a reconstruction error based on the reconstructed image and the source image. The training application updates a first learnable parameter value included in the first CNN based on the reconstruction error to generate at least a partially trained downscaling CNN. The training application updates a second learnable parameter included in the second CNN based on the reconstruction error to generate at least a partially trained upscaling CNN.
What exactly they claim (claim 1)
1. A computer-implemented method for training convolutional neural networks (CNNs) to reduce reconstruction errors, the method comprising: executing a first CNN on a first source image having a first resolution to generate a first downscaled image having a second resolution; executing a second CNN on the first downscaled image to generate a first reconstructed image having the first resolution; computing a first loss based on the first reconstructed image and a second reconstructed image, wherein the second reconstructed image is different from the first reconstructed image, has the first resolution, and is generated by upscaling the first downscaled image using a third CNN; updating a first learnable parameter value included in the first CNN based on the first loss to generate at least a partially trained downscaling CNN; and updating a second learnable parameter value included in the second CNN based on the first loss to generate at least a partially trained upscaling CNN.
Facts
- Company
- Netflix (NETFLIX, INC.)
- Official title
- Techniques for jointly training a downscaler and an upscaler for video streaming
- Patent number
- US12739416B2
- Application number
- 17981281
- Filed
- Nov 4, 2022
- Granted
- Sep 15, 2026
- Inventors
- 2
- Claims
- 20
Classified under
H04N 19/436 Electricity & electronicsG06T 3/4046 Physics & computingH04N 19/30 Electricity & electronicsMore from Netflix in AI & software
- Netflix patents method to bundle game frames for smoother streaming Sep 22, 2026
- Netflix patents search system that explains why a show is missing Sep 29, 2026
- Netflix patents tech to match video scenes across different languages Oct 13, 2026
- Netflix patents method to estimate video depth in real time Sep 24, 2026
- Netflix patents on-device AI to identify objects in streaming video Oct 1, 2026
Summaries are simplified from the filing’s own abstract and first claim. A patent filing is not a product announcement.
Patent data and drawings: U.S. Patent and Trademark Office (USPTO). Nebula is not affiliated with or endorsed by the USPTO.