TS RTX Upscaler

If you have an RTX card, NVIDIA's RTX Video Super Resolution is the fastest honest upscaler in the ecosystem - and the community has settled the question of what it's for. It's not generative. As the KB's upscaling doc frames it (with a +821 thread behind it), NVIDIA's path is "more like Lanczos, an interpolation method" - it cleans up pixelation and makes an already-decent source look sharper, in real time, but it will not invent eyelashes. It's the "more pixels, source already sharp" job, and TS RTX Upscaler brings that onto the canvas for batches of frames.

How it works

The node shells out to NVIDIA's nvvfx Python package (nvidia-vfx on PyPI) and runs its VideoSuperRes engine over your image batch. Quality is picked from the four RTX VSR levels - LOW / MEDIUM / HIGH / ULTRA (the default) - which trade sharpness against VRAM and time. Sizing is either a multiplier (resize_type = "scale by multiplier", scale 1–4, 2 default) or explicit target dimensions (width/height, 64–8192 in steps of 8). Output is a single upscaled_images batch at the resolved size.

The one hard requirement is on the hardware: RTX GPU only. No AMD, no Intel, no Mac - this is NVIDIA's proprietary pipeline, and on anything else the node errors out at import.

The install step that gets everyone

Here's where people actually get burned, and it's worth reading before you install. nvvfx is not in the pack's requirements - the node fails gracefully and tells you what's missing, but you still have to install it yourself:

python -m pip install nvidia-vfx

And here's the trap the community hit: on many Windows setups a plain pip install fails to build, because the package isn't on PyPI proper. The fix people landed on is to install it from NVIDIA's own index:

python -m pip install -U --no-build-isolation nvidia-vfx --index-url https://pypi.nvidia.com

On a Windows portable install, run that from python_embeded\python.exe. And the ComfyUI-native RTX nodes exist separately (Comfy-Org's Nvidia_RTX_Nodes_ComfyUI) - this node wraps the same underlying nvvfx library but lives inside comfyui-timesaver.

cd ComfyUI/custom_nodes
git clone https://github.com/AlexYez/comfyui-timesaver
cd comfyui-timesaver
python -m pip install -r requirements.txt

Common issues

• Node shows "(IMPORT FAILED)" or "nvidia-vfx is not installed." That's the missing dependency - install it with the index-url command above. This is the single most common failure for RTX VSR nodes, and it's exactly what reddit threads kept tripping over.
• "It's fast but it didn't add detail." Correct - that's not its job. If you need detail invented (upscaling a soft face, repairing damage), this isn't the node; the KB's table puts SeedVR2 or a tiled generative upscale in that slot. RTX VSR is for making an already-clean 1080p video look sharp at 4K, fast.
• OOM or quality issues on ULTRA. Drop to HIGH or MEDIUM. ULTRA is a lot of VRAM for marginal returns on most footage.

The honest verdict: for its job - cheap, real-time upscaling of video that's already decent - nothing else in the graph is close on speed. Just don't expect it to resurrect a blurry frame, and budget ten minutes for the nvidia-vfx install dance the first time.