vsdrba-distilled 1.0.0

DistilDRBA function for VapourSynth

VS-DistilDRBA

vapoursynth version for DistilDRBA based on vs-rife.

This project is modified from HolyWu/vs-rife and achieves nearly the same interpolation quality as the original DistilDRBA project.

With TensorRT integration, it achieves a 400% acceleration, enabling real-time playback on high-performance NVIDIA GPUs.

Dependencies

• PyTorch 2.6.0 or later
• VapourSynth R66 or later

trt requires additional packages:

• TensorRT 10.7.0.post1 or later
• Torch-TensorRT 2.6.0 or later

To install the latest stable version of PyTorch, Torch-TensorRT and cupy, run:

pip install -U packaging setuptools wheel
pip install -U torch torchvision torch_tensorrt --index-url https://download.pytorch.org/whl/cu128 --extra-index-url https://pypi.nvidia.com

Installation

pip install -U vsdrba_distilled==1.0.0

Note: Please make sure to install all dependencies listed in the Dependencies section before performing the steps in the Installation section.

If you want to download all models at once, run python -m vsdrba_distilled. If you prefer to only download the model you specified at first run, set auto_download=True in drba_distilled().

Usage

from vsdrba_distilled import drba_distilled
ret = drba_distilled(clip, trt=True, factor_num=2, factor_den=1, scale=1.0, model="v1", auto_download=True)

See __init__.py for the description of the parameters.

Benchmarks

model	scale	os	hardware	arch	speed(fps) 720	speed(fps) 1080	vram 720	vram 1080	backend	verified output	batch	level	streams	threads	trtexec shape	precision	usage
drba_distilled v1	2x	Linux	rtx5070 / 14600kf	drba_distilled	251	115	1.8gb	2.9gb	torch+trt cu128	yes, works	1	5	-	1	static	RGBH	drba_distilled(clip, trt=True, model="v1", trt_optimization_level=5)
drba_distilled v2_lite	2x	Linux	rtx5070 / 14600kf	drba_distilled	999+	700	-	-	torch+trt cu128	yes, works	1	5	-	1	static	RGBH	drba_distilled(clip, trt=True, model="v1", trt_optimization_level=5)

🤗 Acknowledgement

This project is supported by SVFI Development Team.