Vapoursynth GradFun3Mod: Difference between revisions

From wiki.selur.net
Jump to navigation Jump to search
New filter page (via create-page on MediaWiki MCP Server)
 
Update to current Hybrid behaviour; code as <pre> (via update-page on MediaWiki MCP Server)
 
Line 21: Line 21:
| Elast || elast || 1.00 – 10.00 || 3.0 || Width of the soft zone above the threshold in which changes fade out instead of being cut.
| Elast || elast || 1.00 – 10.00 || 3.0 || Width of the soft zone above the threshold in which changes fade out instead of being cut.
|-
|-
| Smooth mode || smode || DFTTest, Bilateral, SmoothGrad multi-stage, DFTTest Mod, Bilateral GPU || Bilateral || Smoothing method. DFTTest Mod is the legacy dfttest variant of GradFun3mod. "Bilateral GPU" behaves like "Bilateral"; whether a GPU is used is chosen in Tools: Bilateral.
| Smooth mode || smode || DFTTest, Bilateral, SmoothGrad multi-stage, DFTTest Mod || Bilateral || Smoothing method. DFTTest Mod is the legacy dfttest variant of GradFun3mod. Whether Bilateral runs on the CPU or a GPU is chosen in Tools: Bilateral.
|-
|-
| Dither mode || mode || ordered (Bayer), round to closest, round, Filter Lite (Sierra-2-4A), Stucki, Atkinson, Floyd-Steinberg, Ostromoukhov, void and cluster || Filter Lite || How the result is brought back to the clip's bit depth.
| Dither mode || mode || ordered (Bayer), round to closest, round, Filter Lite (Sierra-2-4A), Stucki, Atkinson, Floyd-Steinberg, Ostromoukhov, void and cluster || Filter Lite || How the result is brought back to the clip's bit depth.
Line 50: Line 50:
== Example ==
== Example ==
Default settings on a 10-bit YUV 4:2:0 source (640x356):
Default settings on a 10-bit YUV 4:2:0 source (640x356):
<syntaxhighlight lang="python">
<pre>
# debanding using GradFun3
# debanding using GradFun3
clip = deband.GradFun3(clip, tools={'expr': 'akarin', 'dfttest': 'neo_fft', 'bilateral': 'vszip', 'rg': 'zsmooth', 'boxblur': 'vszip'})
clip = deband.GradFun3(clip, tools={'expr': 'akarin', 'dfttest': 'neo_fft', 'bilateral': 'vszip', 'rg': 'zsmooth', 'boxblur': 'vszip'})
</syntaxhighlight>
</pre>


== How Hybrid calls it ==
== How Hybrid calls it ==
* GradFun3Mod works on 8–16-bit integer YUV clips; other formats are converted first.
* GradFun3Mod works on 8–16-bit integer YUV clips; other formats are converted first.
* Threshold and amplitudes are passed as entered; the script treats the threshold as 8-bit scale itself.
* Threshold and amplitudes are passed as entered; the script treats the threshold as 8-bit scale itself.
* With a GPU implementation chosen in Tools: Bilateral, both bilateral entries use the GPU bilateral (same strength as on the CPU); otherwise they use the CPU bilateral.
* With a GPU implementation chosen in Tools: Bilateral, the Bilateral mode uses the GPU bilateral (same strength as on the CPU); otherwise it uses the CPU bilateral.
* Values are only written when they differ from the defaults, which are the script's and fmtconv's own.
* Values are only written when they differ from the defaults, which are the script's and fmtconv's own.
* With a 16-bit clip the script does not dither back, so the dither settings have no effect.
* The dither settings stay enabled; they only have an effect when the script dithers back to a lower bit depth. With a 16-bit clip the script does not dither back, so they have no effect.
* GradFun3Mod also offers the [[Vapoursynth common filter options|common options]] (Masked, ApplyOnlyTo, Merge, OnRegion, Gimmick).
* GradFun3Mod also offers the [[Vapoursynth common filter options|common options]] (Masked, ApplyOnlyTo, Merge, OnRegion, Gimmick).



Latest revision as of 09:50, 16 November 2026

GradFun3Mod removes color banding in flat gradients: it smooths, keeps real detail with an elastic threshold and an edge mask, and dithers the result back to the clip's bit depth.

Filtering → Vapoursynth → DeBand

GradFun3Mod controls in Hybrid
GradFun3Mod controls in Hybrid

What it does

The picture is smoothed (bilateral, DFTTest or multi-stage SmoothGrad). The change is then limited: differences up to Threshold are taken fully, differences up to Threshold × Elast fade out, larger ones keep the source. A detail mask built from the luma protects edges and texture. The script works at 16 bit internally and dithers back to the clip's bit depth at the end; this dithering (error diffusion, ordered or void and cluster, optionally with noise) keeps the gradient smooth after lossy compression.

Settings

Control Parameter Range / values Hybrid default Meaning
Planes planes all, chroma only, Y, U, V, YU, YV all Planes that are filtered; the others are only rounded back without noise.
Threshold thr 0.10 – 10.00 0.35 Strength in 8-bit units at any bit depth; higher is stronger, with more detail loss and edge artifacts.
Radius radius 1 – 100 (2 – 9 with SmoothGrad multi-stage) 12 Size of the smoothing.
Elast elast 1.00 – 10.00 3.0 Width of the soft zone above the threshold in which changes fade out instead of being cut.
Smooth mode smode DFTTest, Bilateral, SmoothGrad multi-stage, DFTTest Mod Bilateral Smoothing method. DFTTest Mod is the legacy dfttest variant of GradFun3mod. Whether Bilateral runs on the CPU or a GPU is chosen in Tools: Bilateral.
Dither mode mode ordered (Bayer), round to closest, round, Filter Lite (Sierra-2-4A), Stucki, Atkinson, Floyd-Steinberg, Ostromoukhov, void and cluster Filter Lite How the result is brought back to the clip's bit depth.
Dither amplitude ampo 0 – 10 1 Strength of the dither pattern.
Noise amplitude ampn 0 – 10 0 Extra noise added while dithering; 0 = off.
Pattern size pat 4, 8, 16, 32 32 Pattern width for void and cluster dithering.
dynamic noise dyn on/off off The ordered dither pattern changes every frame.
static noise staticnoise on/off off The added noise is the same in every frame.
Debug debug on/off off Preview only: shows the detail mask instead of the result.
Tools: Expr – akarin, cranexpr, std akarin Expression plugin; akarin needs AVX2.
Tools: RG – zsmooth, neo_smo zsmooth RemoveGrain implementation.
Tools: DFTTest – neo_fft, dfttest2cpu, dfttest2, vszipcu, vsfeel – DFTTest implementation (DFTTest modes): dfttest2cpu needs AVX2, dfttest2 CUDA (Windows), vszipcu CUDA, vsfeel Vulkan.
Tools: Bilateral – vszip, vszipcl, vszipcu, vsfeel – Bilateral implementation (Bilateral modes): vszip on the CPU, vszipcl OpenCL, vszipcu CUDA, vsfeel Vulkan.
Tools: BoxBlur – vszip, std vszip Box blur implementation (SmoothGrad multi-stage).

Example

Default settings on a 10-bit YUV 4:2:0 source (640x356):

# debanding using GradFun3
clip = deband.GradFun3(clip, tools={'expr': 'akarin', 'dfttest': 'neo_fft', 'bilateral': 'vszip', 'rg': 'zsmooth', 'boxblur': 'vszip'})

How Hybrid calls it

  • GradFun3Mod works on 8–16-bit integer YUV clips; other formats are converted first.
  • Threshold and amplitudes are passed as entered; the script treats the threshold as 8-bit scale itself.
  • With a GPU implementation chosen in Tools: Bilateral, the Bilateral mode uses the GPU bilateral (same strength as on the CPU); otherwise it uses the CPU bilateral.
  • Values are only written when they differ from the defaults, which are the script's and fmtconv's own.
  • The dither settings stay enabled; they only have an effect when the script dithers back to a lower bit depth. With a 16-bit clip the script does not dither back, so they have no effect.
  • GradFun3Mod also offers the common options (Masked, ApplyOnlyTo, Merge, OnRegion, Gimmick).

Background

Origin

GradFun3 comes from Firesledge's Dither package for AviSynth; Gebbi's GradFun3mod added the legacy dfttest mode. The VapourSynth version is based on the ports in fvsfunc and muvsfunc. The resizing options of the original (4:4:4 output, resizer, width/height) are not supported.

Dithering

The dither settings are those of fmtconv's bitdepth function. Pattern size only matters for void and cluster, dynamic noise only for ordered dither. fmtconv also knows a quasirandom mode, which Hybrid does not offer.

Sources