Resize Filters

The resize filters scale the input video frames to an arbitrary new resolution, and can optionally crop the frame before resizing with sub-pixel precision.

The following resizers are included:

  • BicubicResize is similar to BilinearResize, except that instead of a linear filtering function it uses the Mitchell–Netravali two-part cubic. The parameters b and c can be used to adjust the properties of the cubic; they are sometimes referred to as "blurring" and "ringing" respectively.

    If you are enlarging your video, you will get sharper results with BicubicResize than with BilinearResize. However, if you are shrinking it, you may prefer BilinearResize as it performs some antialiasing.

  • BilinearResize uses standard bilinear filtering and is a good choice for smoothing overly sharp sources.

  • BlackmanResize is a modification of LanczosResize that has better control of ringing artifacts for high numbers of taps.

  • GaussResize uses a gaussian resizer, which unlike the bicubics, does not overshoot – but perhaps does not appear as sharp to the eye.

  • LanczosResize is a sharper alternative to BicubicResize. It is NOT suited for low bitrate video; the various Bicubic flavours are much better for this.

  • Lanczos4Resize is a short hand for LanczosResize(taps=4). It produces sharper images than LanczosResize with the default taps=3, especially useful when upsizing a clip.

  • PointResize is the simplest resizer possible. It uses a Point Sampler or Nearest Neighbour algorithm, which usually results in a very "blocky" image. So, in general this filter should only be used, if you intend to have inferior quality, or you need the clear pixel drawings. Useful for magnifying small areas for examination.

  • SincResize uses the truncated sinc function. It is very sharp, but prone to ringing artifacts.

  • Spline16Resize, Spline36Resize and Spline64Resize are three Spline based resizers. They are the (cubic) spline-based resizers from Panorama tools that fit a spline through the sample points and then derives the filter kernel from the resulting blending polynomials. See this thread for the technical details.

    The rationale for Spline is to be as sharp as possible with less ringing artifacts than LanczosResize produces. Spline16Resize uses √16 or 4 sample points, Spline36Resize uses √36 or 6 sample points, etc ... The more sample points used, the more accurate the resampling. Several resizer comparison pages are given in the External Links section.

    • Spline64Resize may be the most accurate of the Resize filters. [Dersch]

    • Spline16Resize is sharper and rings just a bit (which may be desirable with soft sources), and looks pleasing to the eye when enlarging or reducing in moderate amounts. [Doom9]

    • Spline36Resize is somewhere in between the other two.

As with any resampling, there are trade-offs to be considered between preservation (or augmentation) of image detail and possible artifacts (i.e., oversharpening).

Syntax and Parameters

BicubicResize (clip, int target_width, int target_height, float "b", float "c",
               float "src_left", float "src_top", float, "src_width", float "src_height")

BilinearResize (clip, int target_width, int target_height,
                float "src_left", float "src_top", float "src_width", float "src_height")

BlackmanResize (clip, int target_width, int target_height,
                float "src_left", float "src_top", float "src_width", float "src_height", int "taps")

GaussResize (clip, int target_width, int target_height,
             float "src_left", float "src_top", float "src_width", float "src_height", float "p")

LanczosResize (clip, int target_width, int target_height,
               float "src_left", float "src_top", float "src_width", float "src_height", int "taps")

Lanczos4Resize (clip, int target_width, int target_height,
                float "src_left", float "src_top", float "src_width", float "src_height")

PointResize (clip, int target_width, int target_height,
             float "src_left", float "src_top", float "src_width", float "src_height")

SincResize (clip, int target_width, int target_height,
            float "src_left", float "src_top", float "src_width", float "src_height", int "taps")

Spline16Resize (clip, int target_width, int target_height,
                float "src_left", float "src_top", float "src_width", float "src_height")

Spline36Resize (clip, int target_width, int target_height,
                float "src_left", float "src_top", float "src_width", float "src_height")

Spline64Resize (clip, int target_width, int target_height,
                float "src_left", float "src_top", float "src_width", float "src_height")
clip

Source clip; all color formats supported.

target_width, target_height

Width and height of the returned clip.

b, c

Parameters for BicubicResize only.

The default for both b and c is 1/3, which were recommended by Mitchell and Netravali for having the most visually pleasing results.

Set [b + 2c = 1] for the most numerically accurate filter. This gives, for b=0, the maximum value of 0.5 for c, which is the Catmull-Rom spline and a good suggestion for sharpness.

Larger values of b and c can produce interesting op-art effects – for example, try (b=0, c= -5.0).

As c exceeds 0.6, the filter starts to "ring" or overshoot. You won't get true sharpness – what you'll get is exaggerated edges. Negative values for b (although allowed) give undesirable results, so use b=0 for values of c > 0.5.

With (b=0, c=0.75) the filter is the same as VirtualDub's "Precise Bicubic".

BicubicResize may be the most visually pleasing of the Resize filters for downsizing to half-size or less. Doom9 [2]
Try the default setting, (b=0, c=0.75) as above, or (b= -0.5, c=0.25).

Default: 1/3, 1/3

src_left, src_top

See Cropping section below.

Cropping of the left and top edges respectively, in pixels, before resizing.

Default: 0.0, 0.0

src_width, src_height

See Cropping section below.

As with Crop, these arguments have different functionality, depending on their value:

  • If > zero, these set the width and height of the clip before resizing.

  • If <= zero, they set the cropping of the right and bottom edges respectively, before resizing.

Note, there are certain limits:

  • clip.Width must be >= (src_left + width)

  • clip.Width must be > (src_left + right)

  • clip.Height must be >= (src_top + height)

  • clip.Height must be > (src_top + bottom)

...otherwise it would enlarge ("un-crop") the clip, or reduce width or height to 0, which is not allowed.

Default: source width, source height

taps

Parameters for BlackmanResize, LanczosResize, and SincResize only.

Basically, taps affects sharpness. Equal to the number of filter lobes (ignoring mirroring around the origin).

Note: the input argument named taps should really be called "lobes". When discussing resizers, "taps" has a different meaning, as described below:

“So when people talk about Lanczos2, they mean a 2-lobe Lanczos-windowed sinc function. There are actually 4 lobes -- 2 on each side...

For upsampling (making the image larger), the filter is sized such that the entire equation falls across 4 input samples, making it a 4-tap filter. It doesn't matter how big the output image is going to be - it's still just 4 taps. For downsampling (making the image smaller), the equation is sized so it will fall across 4 destination samples, which obviously are spaced at wider intervals than the source samples. So for downsampling by a factor of 2 (making the image half as big), the filter covers 8 input samples, and thus 8 taps. For 3X downsampling, you need 12 taps, and so forth.

The total number of taps you need for downsampling is the downsampling ratio times the number of lobes, times 2. And practically, one needs to round that up to the next even integer. For upsampling, it's always 4 taps.” Don Munsil (avsforum post) | mirror.

Range:

  • 1-100 for BlackmanResize and LanczosResize

  • 1-20 for SincResize

Default:

  • 3 for LanczosResize

  • 4 for BlackmanResize and SincResize

p

Parameter for GaussResize only.

Sharpness. Range from about 1 to 100, with 1 being very blurry and 100 being very sharp.

Default: 30.0

Cropping

  • All resizers have an expanded syntax which crops the frame before resizing:

    BilinearResize(100, 150, src_left=10, src_top=10, src_width=200, src_height=300)

...or more succinctly:

BilinearResize(100, 150, 10, 10, 200, 300)

  • The operations are the same as if you put Crop before the Resize:

    Crop(10, 10, 200, 300).BilinearResize(100, 150)

  • The cropping parameters are all floating point. This allows any Resize filter to be used as a sub-pixel shifter. [IanB]

  • PointResize cannot do subpixel shifting because it uses only integer pixel coordinates.

  • Note that Crop gives a hard boundary, whereas the Resize filters interpolate pixels outside the cropped region – depending on the resizer kernel – bilinear, bicubic etc, and not beyond the edge of the image.

  • As a general rule,

    • Crop any hard borders or noise; Resize cropping may propagate the noise into the output.

    • Use Resize cropping to maintain accurate edge rendering when excising a part of a complete image.

  • Negative cropping is allowed; this results in repeated edge pixels as shown below:

    FFImageSource("resize-sintel-6291.jpg")
BilinearResize(Width, Height, -32, -32, Width, Height)
../../_images/resize-sintel-6291.jpg

Original
../../_images/resize-sintel-6291-shift.jpg

Repeated edge pixels

Examples

  • Cropping:

    Crop(10, 10, 200, 300).BilinearResize(100, 150)

which is nearly the same as:

BilinearResize(100, 150, 10, 10, 200, 300)

  • Load a video file and resize it to 240x180 (from whatever it was before):

    AviSource("video.avi").BilinearResize(240,180)

  • Load a 720x480 (Rec. 601) video and resize it to 352x240 (VCD), preserving the correct aspect ratio:

    AviSource("dv.avi").BilinearResize(352, 240, 8, 0, 704, 480)

which is the same as:

AviSource("dv.avi").BilinearResize(352, 240, 8, 0, -8, -0)

  • Extract the upper-right quadrant of a 320x240 video and zoom it to fill the whole frame:

    BilinearResize(320, 240, 160, 0, 160, 120)

Notes

  • AviSynth has completely separate vertical and horizontal resizers. If input is the same as output on one axis, that resizer will be skipped. The resizer with the smallest downscale ratio is called first; this is done to preserve maximum quality, so the second resizer has the best possible picture to work with. Data storing will have an impact on what mods should be used for sizes when resizing and cropping; see Crop Restrictions.

Changelog

Version

Changes

AviSynth+ r2768

Resizers: don't use crop at special edge cases to avoid inconsistent results across different parameters/colorspaces.

AviSynth+ r2664

AVX2 resizer possible access violation in extreme resizes (e.g. 600->20)

AviSynth+ r2632
Fix: Resizers for 32 bit float rare random garbage on right pixels (simd code NaN issue)
Completely rewritten 16bit and float resizers, much faster (and not only with AVX2)
8 bit resizers: AVX2 support.

AviSynth+ r2487
Added support for RGB48/64, Planar RGB 8/16/Float formats.
Added support for Alpha in planar RGBA and YUVA formats.

AviSynth+ r2290

Added support for 16/32 bit YUV formats (C routine only).

AviSynth+ r1858

Fix: RGB resizers shift horizontally to the opposite direction when src_left param is used.

AviSynth 2.6.0

Added SincResize.

AviSynth 2.5.8

Added BlackmanResize, Spline64Resize.

AviSynth 2.5.6
Added Spline16Resize, Spline36Resize, GaussResize.
Added taps parameter in LanczosResize.
Added offsets in Crop part of xxxResize.

AviSynth 2.5.5

Added Lanczos4Resize.

$Date: 2022/03/07 15:10:22 $