Vapoursynth Resizer
Resizer is the page where the scaling method of the Vapoursynth script is chosen. The target size itself is not set here but on Hybrid's Crop/Resize tab; this page decides how the picture gets there.
Filtering → Vapoursynth → Frame → Resize

What it does
Whenever the output size on the Crop/Resize tab differs from the source, the Vapoursynth script resizes the clip. Without the Resizer checkbox this is a plain one-pass resize with fmtconv and the kernel chosen on the Crop/Resize tab. With the checkbox ticked, the method picked in the list below does the scaling instead; each method has its own page. The Resolution line shows the way the size takes: source size → (lowered size →) target size.
Settings
| Control | Parameter | Range / values | Hybrid default | Meaning |
|---|---|---|---|---|
| Resizer | – | on/off | off | Use the method below instead of the plain resize. Ticking it also switches on resizing on the Crop/Resize tab. |
| Method | – | see below | NNEDI3 | Scaling method; only enabled with Resizer ticked. |
| Lower resolution before resize | – | on/off | off | Shrinks the clip by a factor before the real resize (works with and without Resizer). |
| Factor | – | 1 – 32 | 2 | Divisor for the lowered size (rounded). |
| Only: width / height | – | on/off | off | Lowers only the width or only the height; the two exclude each other. |
| Kernel | kernel | point, rect, linear, cubic, lanczos, blackman, blackmanminlobe, spline36, spline64, spline, gauss, sinc, DPID | lanczos | Kernel for the lowering step; DPID uses the DPID plugin with its defaults. |
| Weighted | – | on/off | off | Blends the result with a softer resize, see Weighted Resize. |
| Stepped Resize | – | on/off | off | Reaches the target in several steps, see Stepped Resize. |
Methods
Only methods whose plugin or script is available are listed; most ML upscalers and DGDecNV are Windows only.
- Classic: NNEDI3 (default), fmtconv, ResampleHQ (linear light), DPID and SSIMD (detail-preserving downscalers), GLSL Resizer (mpv shaders), DGDecNV Crop/Resize (in the decoder).
- ML upscalers: Anime4KCPP, SRMD, RealSR, RealCUGAN, VSGAN, RealESRGAN, BasicVSR++, SwinIR (torch), SwinIR (mlrt), AnimeSR, FeMaSR, GRLIR, RVRT, VSMLRT.
Plain kernels
Without Resizer, the kernel list on the Crop/Resize tab maps to these fmtconv kernels:
| Crop/Resize kernel | fmtconv kernel |
|---|---|
| bilinear | bilinear |
| bicubic | bicubic |
| experimental | spline64 |
| nearest neighbour | point |
| area | box |
| luma bicubic, chroma bilinear | bicubic for luma, bilinear for chroma |
| gauss | gaussian |
| sinc | sinc |
| lanczos | lanczos |
| bicubic spline (default) | spline16 |
| any other | blackman |
Example
Resizer unticked (plain resize with the Crop/Resize tab's default kernel) on a 10-bit YUV 4:2:0 source (640x356), resized to 1280x712:
# Resizing using 10 - bicubic spline
clip = core.fmtc.resample(clip, kernel="spline16", w=1280, h=712, interlaced=False, interlacedd=False) # resolution 1280x712 before YUV420P10 after YUV420P16How Hybrid calls it
- The target width and height come from the Crop/Resize tab. Separately filtered fields get half the target height.
- The plain resize is skipped when the clip already has the target size (not with ML upscalers and not with Lower resolution on).
- fmtconv returns 16-bit for integer input and needs 32-bit float instead of 16-bit float; Hybrid converts and tracks this. With an odd target height, limited-range YUV is converted to full range first. Interlaced content is resized with fmtconv's field handling according to the scan type.
- Lower resolution divides the current size by the factor with the chosen kernel before the real resize.
- After resizing, the pixel aspect ratio of the output is written into the frame properties when it differs from the input.
- The Crop/Resize tab option to downsize 4:2:0 to 4:4:4 is applied when the clip is at least halved in both directions: luma is halved with Spline64 and combined with the chroma.
Background
Kernels
Each classic kernel trades sharpness against ringing: bilinear is soft, bicubic is sharper for upscaling, and gauss does not overshoot at all. fmtconv counts taps as half the number of lobes (default 4) and works at high internal precision.
Sources
- fmtconv documentation
- AviSynth+ resize documentation
- Hybrid's tooltips