2015-07-09, 05:00

I also updated the section in the guide dealing with image doubling:

4K Scaling & Image Doubling

What if you are scaling content to 3840 x 2160 (4K)? The example above assumes all content is being scaled to 1920 x 1080 (1080p). It also assumes most resizing takes place between 720p and 1080p. Let's look at this and three other common resizes comparing the pixel density (increase in pixels) and scaling factors (increase in pixels per inch) involved:

720p -> 1080p

1280 x 720 -> 1920 x 1080

Increase in pixels: 2.25x

Scaling factor: 1.5x

SD -> 1080p

640 x 480 -> 1920 x 1080

Increase in pixels: 6.75x

Scaling factor: 2.25x

1080p -> 2160p

1920 x 1080 -> 3840 x 2160

Increase in pixels: 4x

Scaling factor: 2x

720p -> 2160p

1280 x 720 -> 3840 x 2160

Increase in pixels: 9x

Scaling factor: 3x

Resizing from 720p to 1080p represents a scaling factor 1.5x, while scaling from 1080p to 2160p doubles the resolution by 2x. In situations involving large scaling factors (2x or greater), it may be beneficial to use madVR's image doubling. Image doubling does just that – it takes the full resolution luma and chroma information and scales it by factors of two to reach the desired resolution (2x for a double and 4x for a quadruple). If larger than needed, the result is interpolated down to the target.

Doubling a 720p source to 1080p involves overscaling by 0.5x and downscaling back to the target resolution. Improvements in image quality may go unnoticed in this case. However, image doubling applied to larger resizes of 480p to 1080p or 1080p to 2160p will, in most cases, result in the highest-quality image.

super-xbr vs. NEDI vs. NNEDI3 Image Doubling

Three options are available when image doubling: super-xbr, NEDI & NNEDI3. Unlike linear scalers such as Jinc and Lanczos, image doubling algorithms rely on pattern recognition through statistical sampling. The performance of these algorithms can be slow because several calculations are made to upscale the image resulting in additional GPU processing time. Image doubling algorithms are most effective when applied to resizes at least 2x or larger. Incremental improvement may be observed in smaller upscales, but the corresponding resources consumed upscaling and downscaling may not be worth the extra processing.

super-xbr

Jinc

NEDI

super-xbr

NNEDI3 256 neurons

Image Comparison – Lighthouse:

Jinc

NEDI

super-xbr

NNEDI3 256 neurons

Image Comparison – Lighthouse Top:

Jinc

NEDI

super-xbr

NNEDI3 256 neurons

Note: Chroma upscaling is a form of image doubling, and all three algorithms are available for this purpose. Visual differences between algorithms will be small when upscaling the chroma layer alone.

4K Scaling & Image Doubling

What if you are scaling content to 3840 x 2160 (4K)? The example above assumes all content is being scaled to 1920 x 1080 (1080p). It also assumes most resizing takes place between 720p and 1080p. Let's look at this and three other common resizes comparing the pixel density (increase in pixels) and scaling factors (increase in pixels per inch) involved:

720p -> 1080p

1280 x 720 -> 1920 x 1080

Increase in pixels: 2.25x

Scaling factor: 1.5x

SD -> 1080p

640 x 480 -> 1920 x 1080

Increase in pixels: 6.75x

Scaling factor: 2.25x

1080p -> 2160p

1920 x 1080 -> 3840 x 2160

Increase in pixels: 4x

Scaling factor: 2x

720p -> 2160p

1280 x 720 -> 3840 x 2160

Increase in pixels: 9x

Scaling factor: 3x

Resizing from 720p to 1080p represents a scaling factor 1.5x, while scaling from 1080p to 2160p doubles the resolution by 2x. In situations involving large scaling factors (2x or greater), it may be beneficial to use madVR's image doubling. Image doubling does just that – it takes the full resolution luma and chroma information and scales it by factors of two to reach the desired resolution (2x for a double and 4x for a quadruple). If larger than needed, the result is interpolated down to the target.

Doubling a 720p source to 1080p involves overscaling by 0.5x and downscaling back to the target resolution. Improvements in image quality may go unnoticed in this case. However, image doubling applied to larger resizes of 480p to 1080p or 1080p to 2160p will, in most cases, result in the highest-quality image.

super-xbr vs. NEDI vs. NNEDI3 Image Doubling

Three options are available when image doubling: super-xbr, NEDI & NNEDI3. Unlike linear scalers such as Jinc and Lanczos, image doubling algorithms rely on pattern recognition through statistical sampling. The performance of these algorithms can be slow because several calculations are made to upscale the image resulting in additional GPU processing time. Image doubling algorithms are most effective when applied to resizes at least 2x or larger. Incremental improvement may be observed in smaller upscales, but the corresponding resources consumed upscaling and downscaling may not be worth the extra processing.

super-xbr

- Resolution doubler;

- Relies on RGB inputs - luma and chroma are doubled together;

- Fastest of the three. Slightly faster than Jinc;

- Sharper than Jinc with less ringing;

- Less aliasing on edges than NNEDI3 16 neurons;

- Best bang for the buck.

- Resolution doubler;

- Relies on RGB inputs - luma and chroma are doubled together;

- Second fastest of the three. Slower than super-xbr;

- Least sharp of the three. Best used with SuperRes;

- Known to introduce artifacts when used alone.

- Resolution doubler;

- Uses YCbCr color space - capable of doubling luma and chroma independently;

- Slowest of the three. Slower than NEDI;

- Similar sharpness to super-xbr. But more in-focus;

- Best overall image characteristics - sharpness, aliasing and ringing.

Jinc

NEDI

super-xbr

NNEDI3 256 neurons

Image Comparison – Lighthouse:

Jinc

NEDI

super-xbr

NNEDI3 256 neurons

Image Comparison – Lighthouse Top:

Jinc

NEDI

super-xbr

NNEDI3 256 neurons

Note: Chroma upscaling is a form of image doubling, and all three algorithms are available for this purpose. Visual differences between algorithms will be small when upscaling the chroma layer alone.