Continuity Neural Pass (DLSS 5)

If you've seen Continuity Neural Refine (DLSS 5) and wondered what its sibling does, here's the short version: Refine works on one image, and Pass works on the whole reel. Continuity Neural Pass is the internal, whole-render version of the same NVIDIA DLSS 5 neural renderer - it takes the pack's MMC_REEL (its internal container for the sequence of rendered parts), runs every generated pass through the material refiner in play order, and hands back a refined reel.

Here's the part that will save you confusion: you almost never place this node. When you render with the pack's Creator or Timeline node and the DLSS 5 pill is on, the render loop (render.emit) writes a Neural Pass into your graph for you, after the last pass lands on the reel and after a ReDetail pass if you have one. It's registered as a developer/internal node, so it won't even show up in a plain node search - the pill is the switch you actually touch. See one in a workflow after a render and you're looking at the pack working, not at something you forgot to wire up.

What it does that Refine doesn't

The reel version exists because refining a sequence frame-by-frame is how you get a clip that boils - each frame drifting in its own direction until the video shimmers. Pass instead carries history across the parts in play order, the previous frame's result reprojected into the next through optical flow. That continuity isn't unconditional: if two passes aren't adjacent on the reel - a spliced clip between them is footage nobody refined - history resets and starts fresh, and it also resets at scene cuts. The pack logs how many times that happened, which is a nice honest touch.

It also runs after ReDetail, at whatever size the reel has by then, and keeps that size - where ReDetail changes dimensions, this doesn't. Memory is handled per frame, streamed through a spill/rewrite path rather than held whole, so a ten-second 768p pass costs about what a single frame costs rather than the whole clip. The model graph is loaded once and released between jobs, so it's not squatting on your GPU while a render needs the room.

The dials it has

Compared to the image node, Pass is stripped back on purpose: no processing_scale, no history toggle - it always runs at the reel's own size with history built in. What's left mirrors the Refine node's semantics:

• profile - standard is what the driver runs; natural and cinematic shift the style index; neutral switches local tone and structure off (effectively an off switch).
• detail_strength - how much of the model's high-frequency change is kept; 1 is the model's own answer, default 1.25.
• colour_strength - the low-frequency tone half, defaulting to 0 because at 1 the model's colour behaviour darkens skin and flattens fabric.
• intensity - blend of the refined output over the source.
• precision - reference (float32, matches the driver to ~0.005 MAE) or fast (float16, half the memory).
• frame_index - seeds the deterministic noise for the first frame, and every frame counts up from it. The render's own seed feeds this, so a re-queue gives you the same refined frames and a piece you re-roll gets new noise, the way sampling behaves.

What it needs from you

Exactly what Refine needs, and only once: the DLSS 5 weights don't ship with the pack. You supply your own nvngx_dlssnr.dll (file version 310.8.0.0) from NVIDIA's Streamline SDK or a game that carries it, then on the settings page under Neural refiner press Check and Extract weights. That writes models/dlss/dlssnr-weights-logical.safetensors locally - nothing is downloaded. Install the pack itself the usual way:

cd ComfyUI/custom_nodes
git clone https://github.com/roadmaus/ComfyUI-Continuity

Restart ComfyUI, no pip install needed. If the weights aren't extracted, the pass fails with a plain sentence naming what's missing. If a reel has nothing generated on it, you'll get the "nothing to refine" error - the pass only exists to refine actual generated passes. And remember the pack's rename gotcha: if you already have a ComfyUI-MiniMax-Creator folder, pull inside it instead of cloning a second copy, or two folders register the same node ids and nothing shows up in search.

Want to try DLSS 5 on a single picture without the whole creator? That's the other node. Pass is for whole renders; Refine is for everything else.