FL RTX VSR Runtime Loader

Every workflow in this pack starts here

ComfyUI has no idea what "NVIDIA RTX Video Super Resolution" means out of the box. It's not a model file sitting on Hugging Face - it's a proprietary NVIDIA SDK, and someone has to fetch it and hand it to the graph. That someone is this node. FL RTX VSR Runtime Loader does nothing visible by itself; it exists to emit one thing, a runtime object of the custom RTXVSR_RUNTIME type, which you wire into the runtime input of FL RTX VSR Upscale or FL RTX VSR Restore. No runtime, no processing - the other two nodes will reject a bare image socket if you skip it. Think of it as the power supply for the pack.

It comes from filliptm/ComfyUI-FL-RTXVSR by filliptm (who publishes as Machinedelusions, the author behind the FL- family of packs and a bunch of LTX/LoRA work). The whole family shares one shape: a loader node that resolves heavy machinery, then thin processing nodes that just use it.

What it actually does - and when it does it

Here's the part people find surprising, because it's unlike a model loader you already know. The node holds a pinned NVIDIA package, nvidia-vfx 0.1.0.1 - a ~468 MiB Windows wheel - and on its first execution it downloads that wheel straight from NVIDIA's own package index (pypi.nvidia.com), verifies its SHA-256 checksum, extracts it into ComfyUI/models/rtx_vsr/nvidia-vfx/0.1.0.1, and loads it. Every later run just validates what's already there and skips the network entirely.

The README makes the timing explicit: no network request happens when the loader isn't executed or when a valid runtime is already installed. So you can load the workflow, browse away, and nothing downloads until you actually queue it. If a download dies partway, it resumes from the byte offset via HTTP Range requests rather than starting over, and a checksum failure deletes the partial file rather than trusting it. That's more care than a lot of model downloads get.

The one input and the one output

The schema is almost comically small.

• download_if_missing (boolean, default true) - the author's own tooltip: "Download and verify NVIDIA's pinned nvidia-vfx runtime from NVIDIA's package server when it is missing." Flip it to false and the node insists on a pre-existing, already-valid runtime under models/rtx_vsr. Handy once you've got one installed and want to guarantee an offline workflow never phones home.
• runtime output - wire this into FLRTXVSRUpscale and/or FLRTXVSRRestore. You only need one loader even if both processing nodes are in the graph.

Installing the pack and meeting its real requirements

Install the pack the usual way - in ComfyUI Manager, search ComfyUI-FL-RTXVSR, or:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-RTXVSR.git

Restart ComfyUI. Note that unlike some NVIDIA wrappers, this pack has no requirements.txt to pip-install - the nvidia-vfx runtime is fetched by the loader itself, not by your environment. The actual requirements are platform, and they're firm:

• 64-bit Windows 10 or newer (the code raises on anything else)
• NVIDIA GeForce RTX 20 series or newer (or an RTX professional GPU) - a GTX card will be refused
• A current NVIDIA driver and a CUDA-capable setup
• Python 3.12 or newer (the current ComfyUI portable ships it)
• Roughly 1.5 GiB free disk, since the 468 MiB download needs room to extract

Where people get burned

All the platform checks fire at execution time, not when you add the node. On Linux, Python 3.11, or a non-RTX GPU you'll get a clean error message the moment you queue the workflow - so don't panic-read it as a broken install; read it as "this pack needs that machine." The other classic "is it stuck?" moment is the first run: a ~468 MiB download with a progress bar while your workflow sits there. It's working. Give it a couple of minutes, then the loaders after it are instant.

One more honest expectation: this fetches NVIDIA's proprietary VFX runtime, governed by NVIDIA's license. The pack itself is MIT, but the thing it downloads is not open source. Fine for personal use; read the third-party notice before you build a product on it.