GetFnative

A script that help find the native fractional resolution of upscaled material (mostly anime)

How it works

The general idea can be found in anibin's blog.

getfnative

Suppose

The clip was rendered in a native integer resolution base_width x base_height.
The rendered clip then got upscaled and cropped to only keep the central region (typically 1920 x 1080).
(Optional) Borders of the upscaled clip might be further filled by black (with crop_left, crop_right, crop_top and crop_bottom pixels, respectively) to match a certain aspect ratio (e.g. 1920 x 803).

This script performs a naive search on the fractional src_height corresponding to Step 2 ot the above. The search is taken in the interval from min to max with a given step_length. These parameters can be specified using -min, -max and -sl, respectively.

Since information of the cropped part of the clip is permanently lost, we cannot figure out the exact values of base_height (specified with -bh) and base_width (with -bw). Fortunately, it usually suffices to specify the parity of these values for descaling. Default values are identical to the dimensions of the tested clip. Note that with the wrong parity used you might see a spike in the plot towards an opposite direction.

Pass -m w or -m h to descale only in width or in height. This may be helpful for identifying the parity of base dimensions individually. By default both dimensions are considered.

To acquire the cropping parameters used in descaling, please refer to the function descale_cropping_args.

getfnativeq

Q for quick, as it only checks certain preset values with limited options, such as frame number (-f), base_height (-bh), base_width (-bw) and mode (-m). Each of the guessed values is corresponding to an upscaling ratio of 1.xx or a downscaling ratio of 0.xx, with exactly two decimal numbers.

Examples

python -m getfnative script.vpy -bh 864 -f 1001

This examines frame 1001 from the output index 0 of the provided script, by checking the src_height values of 764.0, 764.25, 764.5, ..., 864.0. By default min = base_height - 100 and sl = 0.25.

python -m getfnative script.vpy -bh 864 -f 1001 -min 820 -max 840 -sl 0.1

The values of src_height to be checked become 820, 820.1, 820.2, ..., 840.

from getfnative import descale_cropping_args
d_args = descale_cropping_args(clip, src_height=830.77, base_height=900, base_width=1600)
descaled = core.descale.Debilinear(clip, **d_args)
f1 = 1
f2 = 1
upscaled = descaled.nnedi3.nnedi3(field=f1, dh=True).std.Transpose().nnedi3.nnedi3(field=f2, dh=True).std.Transpose()
c_args = dict(
    src_width = d_args['src_width'] * 2,
    src_height = d_args['src_height'] * 2,
    src_left = d_args['src_left'] * 2 + 0.5 - f2,
    src_top = d_args['src_top'] * 2 + 0.5 - f1
)
final = core.resize.Bicubic(upscaled, 1920, 1080, **c_args)

A demo script for anti-aliasing, assuming you have got an estimated src_height, say, 830.77, you will descale, upscale with nnedi3, and resize to 1920x1080. These cropping arguments are the ones used for testing the kernel, and is often suitable for post-processing provided the resizer does not have rather large taps.

python -m getfnative script.vpy ... -ct 138 -cb 139

An example for source with letterboxing. -ct and -cb are short for --crop-top and --crop-bottom, respectively. The input clip should have borders already cropped.

from getfnative import descale_cropping_args
# Assuming clip has dimensions 1920 x 803
print(descale_cropping_args(clip, src_height=810, base_height=1080, base_width=1920, crop_top=138, crop_bottom=139))
# Expected output: {'width': 1440, 'height': 603, 'src_width': 1440.0, 'src_left': 0.0, 'src_height': 602.25, 'src_top': 0.5}

Depending on the actual source, both cases may happen:

Upscale a 1440 x 603 clip to 1920 x 804, crop the bottom line, and letterbox;
Upscale a 1440 x 810 clip to 1920 x 1080, and letterbox.

In getfnative the second case is checked, and the result includes a 0.5 pixel shift on top.

Acknowledgment

Thanks to the authors of getnative and Yuuno.

I don't know what I'm doing but you'd better do.