MiniMax H3 DLSS-NR Image / 图片超分 (T8 Advanced)

The name is a bit of a lie, and knowing that saves you an afternoon. "MiniMax H3 DLSS-NR Image Super Resolution" doesn't run any MiniMax model, doesn't call an API, and needs no key. It's a ComfyUI front-end onto NVIDIA's DLSS Denoiser + Super Resolution runtime - a Windows/RTX-only external program you install yourself - used here as a post-process on frames your H3 generations produced. The "T8 Advanced" in the title just means it ships in T8mars/comfyui-minimax-h3-audio-T8, the big single-author H3 workbench where all the serious finishing nodes live.

Why reach for it? H3 video gets soft and noisy at the edges, and if you're pulling stills out for a hero shot or a poster frame you want a clean 2x without firing up a generative upscaler. DLSS-NR is the "more pixels, not invented detail" end of the spectrum - closer to NVIDIA's hardware upscaling path than to a SeedVR2-style restoration model. That's a feature here: it won't rewrite a face or hallucinate pores, but it also won't change identity, lip-sync or any real texture the source already has. The author's own blind tests came back saying no single high-res method beats all others on every source, so think of this as a fast, predictable cleanup, not a magic restore.

How it works

Each SDR image in your batch is pushed independently through the audited external runtime. RGB crosses an explicit 8-bit RGBA bridge; alpha and any extra channels aren't run through the denoiser at all - they're resized from the source and stitched back on, so you don't get corrupted transparency. That's the mechanism most other DLSS wrappers get wrong.

The inputs that matter

You won't set most of these. The node defaults to mode = sr_nr (denoise and upscale), scale = 2.0, and quality_profile = standard, which reproduces the reference ComfyUI-DLSS-NR wrapper's settings. The one non-negotiable input is dlss_nr_runtime - a custom T8_DLSS_NR_RUNTIME handle that can only come from the pack's Runtime Audit node, and only when that audit reports READY. Wire anything else and it refuses, on purpose.

The only other thing a beginner should touch is sr_preset and the named quality_profile options (max_detail, portrait, night, light) if you're denoising faces or dark scenes. Pick custom only if you want to hand-tune nr_intensity, nr_detail, nr_style and friends. Outputs are candidate (the processed image), source (untouched, for A/B), and report_json for the terminal-masochists.

Installing and running it

Install the pack first - ComfyUI Manager, search "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

Then fully restart ComfyUI. No pip packages are involved: the pack deliberately keeps its requirements.txt empty so it can't clobber ComfyUI's Torch stack. What you do need is the external runtime, which the nodes never download for you:

• Windows 10/11 with an NVIDIA RTX GPU and driver 616.56 or newer.
• The full video2dlssnr_release.zip v1.3 from the video2dlssnr release page - not the "light" package.
• Its out directory's four files copied to ComfyUI/models/DLSS-NR/1.3/bin/, plus the pack's runtime manifest saved as models/DLSS-NR/1.3/t8-runtime-manifest.json.

Run the Runtime Audit workflow first; it verifies hashes and the driver, and only a READY result lets this node through.

Where people get burned

• No RTX, no run. On Linux, macOS or non-NVIDIA Windows this is dead on arrival - the audit can't pass without the CUDA/DXGI adapter mapping.
• v1.2 won't do 2x. The v1.2 runtime only allows 1x NR-only; default 2x work requires v1.3.
• The license switch is off by default. accept_external_runtime_license exists because the runtime ships NVIDIA's proprietary DLLs under NVIDIA's terms. You flip it after you've actually accepted those terms yourself.
• Red nodes after install. The pack needs a fairly current ComfyUI with native MiniMax H3 support. Update ComfyUI, the front-end and Manager, then quit and restart fully - updating just the pack usually isn't enough.