ComfyUI's NVIDIA VFX nodes upscale AI video 4x for free

ComfyUI published three NVIDIA VFX nodes on 13 October that upscale, interpolate and convert AI video to HDR inside one node graph. The wrapper code is free. The hardware is not: all three need an NVIDIA RTX card, and one needs a recent one.

The package wraps NVIDIA's RTX Video effects, the same engines behind RTX Video Super Resolution in browsers and players, as nodes that take and return ordinary ComfyUI image batches. The release post lists RTX Video Super Resolution, RTX Video Frame Generation and RTX Video HDR.

What the three nodes do

Super Resolution changes the size of the frame. It scales from 1x to 4x by multiplier, or to target dimensions up to 8,192 pixels, across 19 quality modes for compressed footage, clean sources, streams, denoise and deblur. Output rounds to the nearest multiple of 8.

Frame Generation changes how many frames there are. Multiplier mode runs from 2x to 16x, and timestep mode drops one frame between each pair anywhere from 0.01 to 0.99. For N input frames and a multiplier of M you get (N - 1)M + 1 frames.

HDR changes the brightness and color range. It lifts SDR into BT.2020 with HLG or PQ output, with contrast, saturation, middle gray and peak luminance from 400 to 2,000 nits adjustable, and debanding on by default.

ComfyUI's recommended order is Super Resolution, then Frame Generation, then HDR.

What you need before you install anything

An RTX GPU. The repository is blunt about the rest: the nodes use whichever CUDA device ComfyUI has selected, CPU-only execution is unsupported, non-NVIDIA GPUs are unsupported, and PyTorch must be built for CUDA 13.0 or newer.

Frame Generation narrows that to RTX 40-series and later, so a 3090 can upscale and convert HDR but cannot interpolate.

The three nodes are Apache-2.0. The nvidia-vfx library they call is not. ComfyUI's FAQ says the nodes "are licensed Apache-2.0, so they are free for commercial use," while the library "they depend on is distributed by Nvidia under its own proprietary license." You own the wrapper, not the engine.

How to install it, in order

1. Update ComfyUI first. The release also reworks EXR handling, and previews need the --enable-asset flag, on by default in Comfy Desktop from 1.1.7 and ComfyUI portable from 0.40.0.
2. Open ComfyUI Manager, search for ComfyUI Nvidia VFX Nodes, install, and restart. The manual route is a clone into ComfyUI/custom_nodes and python -m pip install -r requirements.txt with the interpreter that launches ComfyUI.
3. Check that the dependency imports. A ModuleNotFoundError: No module named 'nvvfx' is the most common install failure in the issue tracker, and the README's fix is to reinstall requirements.txt with the right Python.
4. Open an example instead of a blank canvas. The repository ships graphs for image upscaling, video upscaling, frame generation, HDR and a chunked long-video HDR pass, and ComfyUI Wiki reports that version 0.2.1 reached the Comfy Registry on 13 October with an official RTX Video Enhance template.
5. Run one frame before you run a clip. On Linux, at least one user reports the node returning an all-black tensor with no exception raised.

Which node to reach for on a finished AI clip

Start with Super Resolution, and skip Frame Generation unless the delivery needs it.

The mode matters more than the multiplier. Standard modes suppress compression artifacts, and a clean AI render has none, so the suppression only costs detail. ComfyUI's post says the high-bitrate modes exist to upscale "clean or lossless sources without unnecessary artifact suppression," which describes a freshly rendered AI clip. Choose HIGHBITRATE_HIGH or HIGHBITRATE_ULTRA before plain ULTRA.

Reach for DENOISE_* only on a noisy or banded source, and remember those modes ignore the resize settings and return the input dimensions. DEBLUR_* sharpens a soft clip you mean to keep at its current size.

Do the multiplier arithmetic against your delivery target. A 1,280 by 720 clip needs 3x to land at 3,840 by 2,160. Going past 4x is not an option, and 4x on a 1080p source puts you at 7,680 pixels wide.

Add HDR last, and only if something downstream will read it as HDR.

Where it goes wrong: RTX only

AMD, Intel and Apple Silicon users have no path here at all, and the 40-series floor for frame generation cuts out everyone on 30-series cards.

Linux has a second problem. Issue #39 reports Super Resolution returning an all-zero tensor there across every quality level and both resize modes, with no exception raised, and says the same happens outside ComfyUI with the raw library. Issue #18 reports black output on an RTX Pro 6000 with the open driver. Neither is closed.

Where it goes wrong: video memory

The nodes process image batches, and a batch is only as small as the clip you hand it. The pack caps its own working chunk at roughly 16 megapixels of output. That keeps a single 4K pass inside a consumer card. It does nothing for a five-hundred-frame queue.

The open issues show the result. #16 describes system memory filling up around 60 to 70 upscaled images on a 4080 until the machine swaps. #35 reports GPU memory growing per frame and deadlocking on long runs. #31 says any video longer than 10 seconds crashes.

The defense is the chunked pattern ComfyUI already ships for HDR: cut the clip into blocks with Video Temporal Crop, convert each block, join them with Concatenate Video. When one block still runs out of memory, drop the quality mode, the scale or the frames per batch. Comfy's troubleshooting lists the same dials.

Where it goes wrong: frame generation on fast motion

Interpolation invents the frames between two real ones by estimating motion, and estimation fails where the picture stops being predictable. Hands crossing a face, a whip pan, a subject walking behind a pillar, a hard cut: those are where a generated frame shows a smear, a doubled limb or a ghost.

The node detects shot changes automatically and offers a manual shot_change flag for a cut you already know about. That covers the cut case. It does not cover occlusion, and Remio's write-up flags motion errors "especially around occlusion, cuts, and fast movement" as the cost of the technique.

Two settings keep the damage small. Stay at 2x or 4x for frame-rate conversion, because 16x means fifteen of every sixteen frames are guesses. Then check the output frame rate before you encode: multiply the source rate by the multiplier or the clip will run long, which is what the README's timing section covers.

For slow motion, timestep mode is the more careful tool. One frame per pair gives you 2x slow motion without committing the whole clip to a multiplier, and you can inspect each generated frame before it reaches a timeline.

What the release leaves out

NVIDIA has been shipping these effects into ComfyUI since March, when it announced RTX Video Super Resolution as a ComfyUI node at GDC and put the nvidia-vfx wheel on PyPI for free. The repository dates to 9 March under an earlier name, Nvidia RTX Nodes, which is why the issues run back to spring. What is new is the set: frame generation and HDR join the upscaler. Startup Fortune dates the code change to a pull request merged on 7 October, six days before ComfyUI's announcement.

There is no tagged release. GitHub's releases and tags endpoints both return empty, so the only version number is the 0.2.1 in the project file, with no changelog between builds. For a pack a studio might standardize on, that is the gap worth watching.

Install it from ComfyUI Manager by searching ComfyUI Nvidia VFX Nodes, restart, then open the rtx_video_upscale example before you point it at a client clip.