Vapoursynth DPIR Denoise: Difference between revisions
New filter page (via create-page on MediaWiki MCP Server) |
torch filters are Windows only (via update-page on MediaWiki MCP Server) |
||
| Line 10: | Line 10: | ||
== Requirements == | == Requirements == | ||
* Hybrid's torch add-on and an NVIDIA GPU with CUDA. | * Windows, Hybrid's torch add-on and an NVIDIA GPU with CUDA. | ||
* TRT additionally needs Torch-TensorRT from the add-on; it cannot be combined with TensorRT in vs-mlrt filters in the same script. | * TRT additionally needs Torch-TensorRT from the add-on; it cannot be combined with TensorRT in vs-mlrt filters in the same script. | ||
Revision as of 08:02, 15 November 2026
DPIR (torch) removes noise with DRUNet, the denoiser network of DPIR (Deep Plug-and-Play Image Restoration). It runs on an NVIDIA GPU through PyTorch and keeps detail well even at high noise levels.
Filtering → Vapoursynth → DeNoise

What it does
Every frame is passed through DRUNet (the color model for RGB, the gray model for gray clips). Strength tells the network how much noise to expect; higher values smooth more and lose more detail.
Requirements
- Windows, Hybrid's torch add-on and an NVIDIA GPU with CUDA.
- TRT additionally needs Torch-TensorRT from the add-on; it cannot be combined with TensorRT in vs-mlrt filters in the same script.
Settings
| Control | Parameter | Range / values | Hybrid default | Meaning |
|---|---|---|---|---|
| Strength | strength | 0.001 – 1000 | 5 | Noise level the network assumes, on a 0–255 scale; useful values are about 0–50. |
| Tile (w/h) | tile | 0 – 8192 | 0 / 0 | Splits the frame into tiles to save GPU memory; 0 = no tiling. |
| Tile pad | tile_pad | 0 – 64 | 8 | Overlap of the tiles; only with tiling. |
| Streams | num_streams | 1 – 64 | 1 | Number of CUDA streams. |
| Device | device_index | 0 – 4 | 0 | CUDA device. |
| TRT | trt | on/off | off | Uses TensorRT; the engine is built on the first run and cached. |
| FP16 | – | on/off | off | Runs the network in half precision (16-bit float clip); faster and needs less memory. |
Example
Default settings on a 10-bit YUV 4:2:0 source:
# adjusting color space from YUV420P10 to RGBS for vsDPIRDenoise
clip = core.resize.Bicubic(clip=clip, format=vs.RGBS, matrix_in_s="470bg", range_in_s="limited", range_s="full")
# denoising using DPIRDenoise
clip = DPIR(clip, strength=5.000, task="denoise", device_index=0)How Hybrid calls it
- The clip is converted to 32-bit float RGB (gray clips to 32-bit float gray), or to 16-bit float with FP16, and back afterwards.
- Strength, task and device are always written; tiles, tile pad, streams and TRT only when they differ from the defaults.
- With TRT the engines are cached in the engine path from Hybrid's path settings (or the temp folder).
- DPIR also offers the common options (Masked, ApplyOnlyTo, Merge, OnRegion, Gimmick).
- The same network is available without PyTorch as DPIRDenoise (mlrt); for compression artifacts see DPIR Deblock.
Background
DPIR
DPIR (Zhang et al., TPAMI 2021) uses the DRUNet denoiser as a plug-in prior for image restoration tasks such as deblurring and super-resolution. DRUNet was trained on noise levels of about 0–50 and is one of the strongest Gaussian denoisers.
vs-dpir
HolyWu's vs-dpir runs DPIR in VapourSynth with PyTorch, optionally with TensorRT. TensorRT engines are built per model, size, precision and GPU and cannot be moved to another machine.