(Deno) RTX Video Super Resolution

This is the node that gets NVIDIA's RTX Video Super Resolution running inside ComfyUI without you spending an afternoon fighting Python environments. It wraps NVIDIA's official nvvfx.VideoSuperRes - the same VFX filter you'd reach for in a video editor - so you can feed it a frame batch, pick an effect, and get an enhanced batch back. Windows with an RTX GPU is the requirement. If that's you, this is the cheapest real "upscale my video" button in the graph.

What it is and isn't

Before you get excited, know what VSR actually does. NVIDIA's super resolution is a non-generative upscaler - the community's read is that it behaves more like a very good Lanczos than a diffusion model. It adds pixels and cleans up compression artifacts, and on already-clean sources people find its output more "natural" than generative upscalers. What it will not do is invent detail: no pores, no lashes, no fabric weave. If you want a model to hallucinate believable new detail, you want SeedVR2 or SUPIR, not this. If you want bigger, sharper, and fast - this is the one.

The node is intentionally separated from the rest of the Deno pack. It only imports NVIDIA VFX at upscale time, so installing the pack does not drag nvvfx onto your machine. That's deliberate, and it's also why the install story has one extra step (below).

The inputs that matter

• images - your IMAGE batch or video frame batch.
• mode - the operation. This is the heart of it: VSR, High Bitrate, Denoise, Deblur, each with a quality tier. The author's own mode coach is worth trusting: small, low-res, or compressed source → VSR; already clean but you want it bigger and sharper → High Bitrate; noisy or grainy → Denoise; soft or out of focus → Deblur.
• resize_type - Keep Ratio (default), Scale, Manual, Preset Ratio, or Same Size. Pair it with scale (2x default), megapixels, or width/height. divisible_by defaults to 1 so 1920×1080 stays exact.
• device - GPU index, default 0. Only relevant on multi-GPU rigs.

Denoise and Deblur keep the original size and hide the resize controls entirely, which matches how NVIDIA's same-size modes work. The node output is a single images batch you can wire straight into a save or video-combine node.

Installing the RTX part

ComfyUI Manager will install the pack itself, but it deliberately will not auto-install NVIDIA VFX. The flow, straight from the README:

1. Install deno-custom-nodes and start ComfyUI.
2. Add this node and run it once with an image.
3. If NVIDIA VFX is missing, close every ComfyUI window/process.
4. Click the node's How to install button and follow the visual guide: download the ZIP, move it into ComfyUI\custom_nodes\deno-custom-nodes\tools, extract it there, and run install_rtx_vfx.bat.
5. When the BAT asks Install RTX VFX here?, type Y only if the shown path is inside the ComfyUI install you just closed. If it looks wrong, hit N.
6. Wait for INSTALL COMPLETE, restart ComfyUI, run the node again.

For a plain install of the whole pack, the usual works too:

cd ComfyUI/custom_nodes
git clone https://github.com/Deno2026/comfyui-deno-custom-nodes.git
# restart ComfyUI

Where people get burned

The installer refuses to run if ComfyUI is still alive using that Python, and it stops if no NVIDIA GPU is detected - so "nothing happens" usually means one of those, or that you never closed ComfyUI first. The subtler trap: if another RTX/Broadcast node already loaded NVIDIA's native DLLs, the node will stop and ask for a full restart rather than fight over the native runtime. If another Broadcast-based RTX node works on your box but this one doesn't, that's a native runtime conflict, not a broken install. Restart everything and try again before you suspect the pack.