Vapoursynth VSMLRT Resize

From wiki.selur.net
Revision as of 09:53, 16 November 2026 by Selur (talk | contribs) (Update to current Hybrid behaviour; code as <pre> (via update-page on MediaWiki MCP Server))
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

VSMLRT runs any ONNX super-resolution model you supply (for example ESRGAN-type "2x_…" or "4x_…" models) through vs-mlrt, on the CPU or on GPU backends. It is one of the resizers of Vapoursynth Resizer; several models can be chained, and an optional color fix matches the colors back to the source.

Filtering → Vapoursynth → Frame → Resize

VSMLRT controls in Hybrid
VSMLRT controls in Hybrid

What it does

Every frame is passed through the selected model(s). Hybrid takes the upscale factor from the file name and resizes the result to the exact output size afterwards. Without a model the network step is skipped and the clip is only resized.

Requirements

  • Hybrid's vs-mlrt add-on with at least one backend, and one or more ONNX model files.
  • GPU backends: TensorRT and ONNX (CUDA) for NVIDIA, TensorRT RTX for RTX cards, OpenVINO (GPU) for Intel, NCNN for Vulkan GPUs, DML for DirectX 12 GPUs (Windows); on macOS ONNX (CoreML).

Settings

Control Parameter Range / values Hybrid default Meaning
Model / Select network_path ONNX file(s) empty Model to run; the factor is the number before "x_" in the file name.
Multi – on/off off Select several models, applied in order.
No adjust – on/off off With Multi: do not resize back to the input size between the models.
Serialize path_is_serialization on/off off Loads the ONNX file in the script and hands the model to vs-mlrt in memory instead of as a path; not for TensorRT and TensorRT RTX.
Backend backend OpenVINO (CPU/GPU), ONNX (CPU/CUDA), DML (GPU), TensorRT, TensorRT RTX, NCNN (macOS: ONNX CoreML/CPU) OpenVINO (CPU) (macOS: CoreML) Runtime that executes the network.
32bit fp16 on/off on Full precision; off runs in half precision. Forced on for OpenVINO and TensorRT RTX.
BF16 bf16 on/off off bfloat16; TensorRT and OpenVINO (CPU) only. Switches 32bit on.
Device device_id 0 – 4 0 GPU to use; not for CPU and CoreML backends.
Streams num_streams 1 – 16 1 Parallel streams; TensorRT, TensorRT RTX, OpenVINO (GPU), NCNN and CoreML only.
Cuda Graph use_cuda_graph on/off off TensorRT and ONNX (CUDA) only.
TRT backend optimization builder_optimization_level 0 – 5 3 TensorRT engine build effort.
TRT max workspace workspace 0 – 1048576 MiB 0 TensorRT workspace limit; 0 = not set.
Tile (wxh) tilesize 1 – 1024 off, 256×256 Splits the frame to save memory; off = whole frame.
Overlap overlap 2 – 624 8 Overlap of the tiles; only with tiling.
MultiInput / count – on/off, 2 – 50 off, 2 For models that take several consecutive frames: feeds that many neighbor frames.
Requires: Luma scale / Matrix / Modulo – no, pc, tv / matrix list / 1 – 128 pc / no / 1 Overrides the range, matrix and size multiple Hybrid prepares for the model.
Dynamic Shapes (opt/min/max) opt_shapes, min_shapes, max_shapes 0 – 8192 unset TensorRT engine sizes; with a minimum size the engine becomes dynamic.
Fix Color: Radius, Plane, Mode, Wavelets, Fast – radius 1 – 100, average/wavelet, wavelets 1 – 15 off (radius 10, wavelets 5) Matches the colors of the output to the clip before the model.
Resizer – point, rect, linear, cubic, lanczos, blackman, blackmanminlobe, spline36, spline64, spline144, spline, gauss, sinc spline64 Kernel that brings the result to the exact output size.

Example

A 2x ONNX model on a 10-bit YUV 4:2:0 source (640x356), resized to 1280x712 (default backend, no tiling):

clip = core.std.AddBorders(clip=clip, left=0, right=0, top=2, bottom=2) # add borders to achieve mod 8 (vsVSMLRT) - 640x360
# changing range from limited to full range for vsVSMLRT
clip = core.resize.Bicubic(clip, format=vs.YUV420P10, range_in_s="limited", range_s="full")
# setting color range to PC (full) range.
prop_name = '_Range' if core.core_version.release_major >= 74 else '_ColorRange'
clip = core.std.SetFrameProps(clip=clip, **{prop_name: vs.RANGE_FULL})
# adjusting color space from YUV420P10 to RGBS for vsVSMLRT
clip = core.resize.Bicubic(clip=clip, format=vs.RGBS, matrix_in_s="470bg", range_in_s="full", range_s="full")
# resizing using VSMLRT, target: 1280x712
from vsmlrt import Backend
clip = vsmlrt.inference([clip], network_path="F:/Hybrid/64bit/onnx_models/2x_AnimeJaNai_Standard_V1_Compact_net_g_120000.onnx", backend=Backend.OV_CPU(fp16=False,bf16=False))
clip = core.std.Crop(clip=clip, left=0, right=0, top=4, bottom=4) # removing added borders from mod requirement (vsVSMLRT) -  1280x712
# making sure 0-1 limits are respected
clip = core.vszip.Limiter(clip=clip, min=[0,0,0], max=[1,1,1])

How Hybrid calls it

  • Selected with Resize enabled and Resizer set to VSMLRT (see Vapoursynth Resizer).
  • The clip is converted to 32-bit float RGB (16-bit float with 32bit off), full range, and brought to mod 8 (mod 1 for TensorRT) unless the Requires settings say otherwise.
  • The factor comes from the file name: "4x_name.onnx" counts as 4×; a prefix that is not a number is assumed to be 4×; a file without "x_" is skipped with a comment in the script.
  • After each model the size is multiplied by its factor. With Multi, Hybrid resizes back to the input size between the models unless No adjust is set. At the end fmtconv resizes the result with the Resizer kernel to the exact size if needed; with separately filtered fields the target height is halved. The color fix runs last.
  • With tiling the overlap is always written (vs-mlrt's own default would be 0).
  • With MultiInput the model gets the requested number of neighbor frames; at the start and end of the clip the first or last frame is repeated.
  • TensorRT builds an engine on the first run for each size. TensorRT cannot be used in vs-mlrt filters and torch filters in the same script.
  • The PyTorch counterpart for user-supplied models is VSGAN.

Background

vs-mlrt

vs-mlrt runs ONNX models in VapourSynth with several backends (OpenVINO, ONNX Runtime, TensorRT, NCNN and others) behind one Python wrapper. Its generic inference function takes any ONNX model and one or more input clips; by default it processes the whole frame and uses no tile overlap.

Sources