DLSS SR + Experimental Neural Rendering (Advanced)

This is the pack's workhorse: DLSS Super Resolution followed by the experimental neural-rendering pass, in one node, with the temporal controls the Easy pipeline keeps hidden. It's the "Advanced" node for a reason - it assumes you've already built the guide chain (depth node → maybe a stabilizer, motion node) and just want the native runtime stage done right.

You feed it color, depth, and motion vectors for every frame. It upscales with DLSS SR, runs the neural pass on the result, and stitches the whole clip together. Where it beats the Easy node is video: this is where you decide how long clips get processed, and that decision is most of what determines whether your run finishes or dies an out-of-memory death.

Processing mode: the decision that matters

• Persistent full sequence - one native SR context and one NR context live for the entire batch. No chunk resets, no crossfade seams, and in the author's own 33-frame comparison it was the fastest mode. The trap: ComfyUI still holds the entire input and output batch in memory, so "persistent" is really "as long as your RAM and temp disk hold out." The README measured a 241-frame 1080p clip fine on an RTX 5090 box with 32 GB-plus; multi-hour films are explicitly off the table until a file-to-file streaming node exists.
• Bounded overlap-add - processes overlapping native windows and blends the duplicate frames with a raised-cosine curve. Slower, more RAM-friendly, and the recommended fallback. Start with chunk_size 8 and history_overlap 8, exactly as the docs suggest. Bigger chunks cut overhead; more overlap gives each fresh native context more time to settle in exchange for repeated work.

Watch the runtime report for the mode actually used per window. Ghosting around the seams usually means your overlap is too small or you're in persistent mode with a hard scene cut mid-clip.

The rest of the knobs, briefly

The style group - style (0 neutral, 1, 2), style_strength, intensity, local_structure (0–2), skin_structure (-1 = leave the runtime's behavior alone) - plus auto_mask (the runtime's automatic effect masking, on by default) all get forwarded to the neural pass. sr_quality has the full six choices including Ultra Quality and DLAA; scale sets 2x/3x/4x. Quality presets are subject to whatever your runtime DLL actually supports, so don't panic if Ultra Quality behaves like Balanced.

depth_inverted (off by default) exists because DLSS expects a specific depth convention and your guide node might not match it. If the neural pass produces garbage that looks like near/far confusion, flip it. If the still or clip looks subtly wrong in structure but not inverted, it's more likely your depth or motion guide quality, not this toggle.

Guides must match the batch: image, depth, and motion_vectors all need the same frame count, and the motion must use the pack's current-to-previous encoding (0.5 = zero motion) that the RAFT and Optical Flow nodes emit. Plug in vectors from some other optical-flow pack with a different convention and you'll get exactly the kind of smearing you'd blame the runtime for.

Realistic setup

Run it first with DLSS Runtime Setup (One Click) done, runtime status READY, and weights downloaded (the guide models fetch on first use - the first run can sit there looking dead). On a still image the neural runtime duplicates the frame internally to initialize its temporal state, so don't read quality into a single-frame test - video is the point of this node. Install the pack from https://github.com/HECer/ComfyUI-DLSS5 via ComfyUI Manager, and start with short clips and Bounded overlap-add before you commit to a long persistent run.