MiniMax H3 DLSS-NR Runtime Audit (T8 Advanced)

Every DLSS-NR node in the T8mars H3 pack demands a T8_DLSS_NR_RUNTIME handle on its first input, and this audit node is the only place that handle is born. If you've ever loaded a workflow, watched it light up red, and hunted for a "runtime" node that didn't exist yet - this is that node. Run it first, wire its output everywhere, and the rest of the chain stops complaining.

The setup problem it solves is real and specific. DLSS-NR isn't a PyTorch model ComfyUI can pull from HuggingFace. It's an external Windows program - video2dlssnr.exe plus a stack of NVIDIA DLLs - that you have to source yourself under NVIDIA's license, and the pack refuses to auto-download, install, or redistribute any of it. Given that you're hand-placing proprietary binaries, someone has to check you did it right before any of your pixels go near it. That someone is this node.

What it actually checks

The audit is deliberately read-only. It never downloads anything, never upgrades a driver, never touches your media, and in its default static_only mode never even invokes the runtime. What it does instead:

• confirms the allowlisted official v1.3 release archive exists and matches, under ComfyUI/models/DLSS-NR/1.3/
• verifies the byte identity of the installed EXE and DLLs in bin/
• checks you're on NVIDIA driver 616.56 or newer
• maps the CUDA device index to the DXGI adapter so the runtime talks to the GPU you think it's talking to
• confirms the GPU has enough free VRAM to survive the run - the pack's floor is roughly 512MiB free, and the report surfaces the actual figure

That byte-identity check matters more than it sounds. DLSS-NR has "light" builds, old builds, and DLLs from unrelated NVIDIA packages floating around, and mismatched versions fail in confusing ways mid-upscale. Hashing first means the failure happens here, in one clear report, instead of 400 frames deep.

The inputs you'll touch

• runtime_version - defaults to 1.3, which you want for anything 2x.
• accept_external_runtime_license - starts false, on purpose. You flip it only after you've obtained the runtime yourself and actually read the NVIDIA and upstream terms. The workflow's license gate being off by default is the author's way of not clicking "I agree" on your behalf.
• probe_mode - static_only is the safe default. feature_probe_1_frame goes further and runs a real feature probe on your GPU; the audit docs describe it as the genuine check that the route works on your hardware.

The advanced dxgi_adapter_index and cuda_device_index inputs only matter on multi-GPU boxes; leave them at 0 otherwise.

Outputs are dlss_nr_runtime (the handle), ready (the boolean you check), status, and report_json.

Installing and running it

Install the pack via ComfyUI Manager ("MiniMax H3 Audio T8") or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

Restart ComfyUI completely. Then get the runtime yourself - Windows RTX + driver 616.56+, the full video2dlssnr_release.zip v1.3 from the official release, its four out files into ComfyUI/models/DLSS-NR/1.3/bin/, and the pack's manifest saved as models/DLSS-NR/1.3/t8-runtime-manifest.json. There are no new pip dependencies for this node - it leans on ComfyUI's existing Torch/NumPy/PyAV.

Gotchas

• ready = false is the expected outcome on first run. Read the status output - it tells you exactly which check failed. It's usually the missing manifest, the light package, or an old driver.
• On Linux or macOS this node can never pass. It's Windows/RTX-only by nature of what DLSS-NR is. Don't fight it.
• No dlss_nr_runtime output means you clicked Queue while it was still red - check the audit node first, not the upscaler.