NVIDIA DLSS Super Resolution (Unofficial Bridge)

Most "upscale in ComfyUI" nodes wrap a diffusion model. This one wraps the actual thing - the real NVIDIA DLSS Super Resolution runtime, the upscaler from games, driven through a VapourSynth bridge by VapourKit. It's not an NVIDIA product and it's not affiliated with them; it's an unofficial bridge that takes your image or video batch, hands it to the genuine DLSS SR DLL, and returns the result. The README's test sheets show it doing what DLSS does: doubling a low-poly game-style still from 1672×941 to 3344×1882, sharpening edges and cleaning up blockiness without changing the framing.

The "Unofficial Bridge" in its display name isn't decoration. You get this working by installing the runtime (the VapourKit build supplies the SR wrapper and nvngx_dlss.dll), and the whole chain only runs on Windows with an NVIDIA GPU and a recent driver.

What DLSS SR actually is - so you expect the right thing

This matters more than any setting, because it's where people get burned. DLSS Super Resolution is a temporal neural upscaler: it reconstructs a higher-resolution frame from a lower-resolution one using motion and depth history to reuse detail across frames. It is not a generative restorer. The community's own upscaling wisdom draws exactly this line - NVIDIA's hardware path is read as "more natural" and closer to an interpolator than a diffusion model on already-clean sources, and it will not invent pores, lashes, or fabric weave the way SeedVR2 or SUPIR can. Use it when your source is decent and you want more pixels with temporal stability. Use a generative restorer when your source is soft or damaged and you want hallucinated detail.

The flip side of that honesty: this is the only upscaler in this pack that needs both a depth map and a motion field as required inputs, because that's what lets it do the temporal reconstruction. In a game the engine supplies those. Here, you wire up the pack's guide nodes - Depth Anything V2 or Video Depth Anything for depth, RAFT or Optical Flow for motion_vectors. For a single still, motion is trivially zero and the runtime duplicates the frame internally to seed its temporal state.

The inputs

• scale - 2x, 3x, or 4x.
• quality - the full six: Quality, Balanced, Performance, Ultra Performance, Ultra Quality, DLAA. The node passes these through as integer presets to the runtime, so which ones behave differently depends on your specific DLL. "Subject to runtime support" is the author's own caveat, and it's the correct one.

depth and motion_vectors must match the image batch count, and both are resized to the source internally - so you can feed full-res guides and let the node align them. Connect one guide and not the other and the node refuses, because a half guide defeats the whole point.

Outputs and workflow position

You get image at the new resolution and runtime_report - a STRING naming the official SR chain, the exact before/after dimensions, scale, and quality preset. It's a nice provenance record if you're A/B testing presets on the same clip.

For the guided path, run this node when you want SR only. The Full Pipeline node wraps it plus the neural-render pass when you want both. Before anything runs, the ritual: install the pack via ComfyUI Manager (https://github.com/HECer/ComfyUI-DLSS5), run DLSS Runtime Setup (One Click), and confirm DLSS 5 Runtime Status reads READY - this node specifically needs the SR half of the runtime present. Start with 2x and Quality on a short clip, read the runtime report, and remember that on this pack, the upscale is the reliable part; the experimental neural rendering next door is where the wild stuff happens.