Official Topaz · 环境检查 (T8 EXP)

Topaz is still the paid default everyone benchmarks against - over 200 threads in the 2026 H1 corpus alone - and if you already own a licence, calling tvai_up from inside ComfyUI is genuinely useful. The annoying part was never the call. It was working out which of the three Topaz directories each thing lives in, and whether the 4x weights you want are even on disk.

This node answers both. It's the environment check you run once, before the video node, and it's mandatory: the enhancement node takes its output as an input.

What it actually does

You give it three strings - install_directory, model_definitions_directory, model_data_directory - and it audits them. On the author's machine, to show how much these diverge, they're G:/Program Files/Topaz Labs LLC/Topaz Video for the program, C:/ProgramData/Topaz Labs LLC/Topaz Video/models for the definitions, and G:/ProgramData/Topaz Labs LLC/Topaz Video/models for the data. Yours will be different, and yes, the last two can differ from each other. Fill in real paths; leaving any of them blank throws immediately, which is the correct behaviour for a node whose entire job is paths.

Outputs are topaz_runtime, a custom T8_OFFICIAL_TOPAZ_RUNTIME handle you wire into the video node, and report_json.

The report is the point

Open report_json - that's where the value is. It carries a model_catalog: the model IDs from the definitions directory, which scale weights actually exist for each one, and the parameters that model and version really support, with ranges and defaults. It also separates three categories you must not mix up: plain enhancement definitions, models that need an official Neuroserver, and frame-interpolation/auxiliary models. Only the first group belongs in the enhancement node.

Scale availability is reported honestly too. Something says "candidate weights missing" or "candidate file present, not validated for inference" - so the report won't let you believe Iris3 4x is ready when only 1x and 2x have candidate files. On the author's install it recognised 89 definitions: 52 plain enhancement, 6 Neuroserver, 31 interpolation/auxiliary or unclassified, plus 6 non-model JSON files. That's not 89 runnable models, and the report doesn't claim it is.

It stays read-only on purpose: it reads directories and small definition files, doesn't walk licence data, doesn't hash multi-gigabyte weights, and doesn't execute an enhancement or load Torch. The real weight identity gets checked at execution time instead.

One thing to be blunt about

The Starlight ("星光") models need a separate official Neuroserver with valid entitlement. On the author's machine the probe returned 15 / "License checkout failed", and the whole Starlight line was then parked at the user's request. If you were hoping this pack gives you a cheap route to Topaz's newer models, it doesn't - the node reports what your licensed install can do and nothing more. Same for installing: this node does not install, download or accept agreements for anything. Use the official Topaz program normally, and if you don't have Topaz at all, every other node in this pack keeps working.

Install

Manager → search MiniMax H3 Audio T8, or clone it and restart:

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

No extra Python packages. Topaz is closed-source commercial software, so the license you accepted with the installer governs what you're allowed to do here - running it headlessly on your own footage is your call to make, not the pack's.