VapourSynth-ILS

CUDA implementation of Real-time Image Smoothing via Iterative Least Squares for VapourSynth.

It is a global optimization based edge-preserving smoothing filter, which can avoid haloing and gradient reversal artifacts commonly found in weighted average based methods like bilateral filter and guided filter.

Requirements

• CUDA-enabled GPU(s) of compute capability 5.0 or higher (Maxwell+).

• GPU driver 450 or newer.

• cuFFT library, i.e. cufft64_10.dll on Windows or libcufft.so.10 on Linux.

Parameters

ils.ILS(clip clip[, float lambda=0.5, int iteration=4, float p=0.8, float eps=0.0001, float gamma=None, float c=None, bool use_welsch=False, int device_id=0, int num_streams=2, bool use_cuda_graph=True])

In short, use use_welsch=True with lambda, iterations, gamma for compression artifacts removal tasks or use_welsch=False with lambda, iterations, p for detail manipulation tasks.

• clip

  The input clip. Must be of 32 bit float format. Only the first plane is processed.

• lambda

  Smoothing strength of the filter.

  Default: 0.5

• iteration

  Iteration number of optimization. A larger iteration number can lead to stronger smoothing on large-amplitude details at the expense of a much higher computational cost.

  Default: 4

• p

  Power norm of the penalty on gradient, which controls the sensitivity to the edges in the input image. A smaller value tends to blur smooth regions but leaving salient edges untouched. A value in 0.8 ∼ 1 may be suitable for tasks of tone and detail manipulation, which can produce results with little visible artifacts.

  Default: 0.8

• eps

  Small constant to make the penalty function differentiable at the origin. A larger leads to higher convergency speed with the risk of resulting in halo artifacts.

  Default: 0.0001

• gamma, c

  Computed automatically.

  Default:

  • gamma: 0.5 * p - 1

  • c: p * (eps ** gamma)

• use_welsch:

  Whether to use the Welsch penalty function. If not, the Charbonnier penalty is used instead.

  The Welsch penalty is suitable for clip-art compression artifacts removal while the Charbonnier penalty is suitable for tone and detail manipulation.

  Default: False

• device_id

  Set GPU to be used.

  Default: 0

• num_streams

  Number of CUDA streams, enables concurrent kernel execution and data transfer.

  Default: 4

• use_cuda_graph

  Whether to use CUDA Graphs to reduce CPU cost and kernel launch overhead.

  Default: True

Compilation

cmake -S . -B build -D CMAKE_BUILD_TYPE=Release -D CMAKE_CUDA_FLAGS="--threads 0 --use_fast_math -Wno-deprecated-gpu-targets" -D CMAKE_CUDA_ARCHITECTURES="50;61-real;75-real;86"

cmake --build build --config Release