avisynth: what is the internal faster resize algorithm?

[marcorocchini]

Hi cats please can you tell me the more faster resizer? quality is not important

[sneaker]

PointResize is the fastest but quality is horrible. Bilinear is better but still fast (though a bit blurry). Bicubic is good quality/speed compromise, esp. when downsizing.

[marcorocchini]

PointResize is the fastest but quality is horrible.

^^ thanks

[Bernix]

Hi,
I know you dont mind quality, and sneaker noticed it is horrible. If really it can be called as resizer.
Check this page to imagine it. http://hermidownloads.craqstar.de/videoresizefiltercomparasion/

Bernix

[jagabo]

Check this page to imagine it. http://hermidownloads.craqstar.de/videoresizefiltercomparasion/

Note that all of the images on that page were the result of eight successive resizes, upscaling by 32 pixels each time. For example:

# 256x256 source image
PointResize(width+32, height+32)
PointResize(width+32, height+32)
PointResize(width+32, height+32)
PointResize(width+32, height+32)
PointResize(width+32, height+32)
PointResize(width+32, height+32)
PointResize(width+32, height+32)
PointResize(width+32, height+32)
# 512x512 after resizing

A single resize from 256x256 to 512x512 with PointResize() would simply double the width and height of each pixel.

[Bernix]

Hi,
so you are saying that it is meaningless comparasion ?
Thank you

A single resize from 256x256 to 512x512 with PointResize() would simply double the width and height of each pixel.

I see your point now.

But am I right, that works when resizing by 2x, 4x, etc. Multiplying by some fraction will look horrible.

Bernix

[jagabo]

I use PointResize() as a special purpose filter. It's great for examining individual pixels, zooming by 4x, 8x, 16x... so you see each pixel of the source as a big square block. Or when you need to losslessly upscale by 2x and the downscale back to 1x (when deblocking interlaced video, for example). I pointed it out earlier simply because the image at the web site was especially bad. But all the resizers generate different types of artifacts with incremental zooming than with s single zoom.

Sharper isn't always better. If a video is already sharp a sharpening resizer like Lancoz3 will generate oversharpening halos. Those halos may be further exacerbated if the video is scaled again while viewing (modern TVs usually use sharpening resizers to upscale). Also, any block artifacts in your source will be accentuated by a sharpening resizer. For downscaling, bilinear (AviSynth's BilinearResize() or VirtualDub's Precise Bilinear) is the most natural. It looks most like what you would see with your own eyes viewing the original image from twice as far away. I usually test a few resizers on the video I'm scaling, and use whatever looks best to me.

I'm not sure the use of that particular zone plate is a good way to test resizing algorithms.