DLSS 5 in ComfyUI: we are publishing our pack, and what we have learned
10 octobre 2026 · Kris Wilhem
NVIDIA presented DLSS 5 Neural Rendering as a leap in realism for video games: skin, materials, light. We ran it on something other than games, on images and video shots, in ComfyUI. Today we are publishing the tool we drew from it: a free ComfyUI pack, under the MIT licence, with its nodes, its workflows and what our measurements taught us.
What it changes on a face
Drag the slider: on the left the original image, on the right the same one after DLSS 5, with the setting shipped in the pack.
And on a generated landscape
And on a video shot
The two videos play together: drag the slider during playback.
What we have learned
- The setting makes the render. We first attributed our best results to a depth map. The same setting, with no depth at all, gives the same render: to the eye, no difference.
- The engine ignores depth. Flat, inverted or replaced with noise: the image is identical to the pixel. DLSS 5 reads the image and the motion, it infers the volumes and gives them light.
- It transforms more than it improves. It pulls the image towards its own realism: spectacular on a smooth image, generated or drawn; to be watched on an already photographic portrait, where it guesses the materials.
- In video, time serves stability, not detail. On real shots: no added flicker, brightness preserved.
- There is no universal setting. Skin, stone and metal do not call for the same work. The shipped setting is deliberately light: a starting point.
What the pack contains
- A DLSS 5 render for one image or a batch, starting from the setting we validated, with the native DLSS controls (style, intensity, local light, material, skin);
- a colour control mask: red for intensity, green for local light, blue for material, zone by zone;
- an explorer that outputs every variant of an image on a single contact sheet, each with its settings written beneath;
- the DLSS 5 finish for a video shot;
- a before / after slider comparator, for image and video, inside ComfyUI;
- three ready-to-use workflows and examples.
The pack calls the community engine, Merserk's Visual Enhancer (version 15); it does not contain it. DLSS 5 Neural Rendering runs on a proprietary, unmodified NVIDIA library that NVIDIA does not distribute outside games: our repository contains none of it, nor a single line of NVIDIA code. A recent NVIDIA RTX card is required.
The repository: github.com/Axiris/ComfyUI-AXIRIS-DLSS5
A guide will follow: how to tune the workflow image by image, with more examples. At Axiris, DLSS 5 has become a finishing pass on our video shots and a tool of our studio.