Official Topaz · 视频高清增强 (T8 EXP)

Despite living in the audio-centric T8 pack, this node never touches your sound. It enhances picture through Topaz's official tvai_up filter, and it then goes out of its way to prove your audio came through untouched - packet count, decoded PCM, start offset, the lot.

That's the reason to care. Community upscalers like SeedVR2 and FlashVSR are excellent and free, and one corpus post even reports dropping a Topaz subscription in favour of an open node. What they don't give you is Topaz's delivery hygiene: this node emits a lossless MOV or MKV master, then verifies the frame timeline, the audio packets and the decoded PCM rather than just checking the file exists.

How it works

It needs topaz_runtime - the handle from the environment node, not a path you type here. It takes source_video as a VIDEO input (from a real file: if you cropped or trimmed in the graph, save that first, because the node works on the file), runs tvai_up out-of-process, and writes the master.

Container choice is driven by your audio. Videos with AAC get PNG48 lossless MOV, other supported tracks get FFV1 MKV, and the audio is never transcoded. That's a fix for a real bug class: MP4 AAC padding metadata getting lost when you remux to MKV, which is the kind of thing that shows up as a half-frame of silence at the front of your clip.

model_id is the definitions filename without .json - iris-3, not "Iris". scale is 1x (same size, enhancement only), 2x or 4x, and it does not change frame rate; frame interpolation is a different family of models and isn't this node's job. vram_fraction (0.8) is the share of VRAM the external engine may use, not your total. parameters_json defaults to {} and otherwise takes only what your chosen model and version genuinely support - {"noise":0.3,"details":0.5} is the author's example, and an unsupported key will fail rather than being ignored.

The optional size_mode lets you go off-scale: keep the default scale, or switch to target_dimensions and fill target_width/target_height (even numbers, 32–8192, same aspect ratio, 1–4x). Then scale is ignored. Handy when you want a specific delivery size out of a 1024×512 source.

Outputs: enhanced_video, source_video passed through, saved_path, and report_json.

What's actually verified, and what isn't

Verified: an Iris2x public-node run took 1024×512 to 2048×1024, 72 frames at 24fps, with all 96 original audio packets, the AAC padding, the decoded PCM and the audio/video start offset preserved. A 24-second test hit 2048×1024, 576 frames, 752 audio packets, same story. Also useful: 1x on a 1024×512/72-frame MP4 produced 95 packets with PCM and relative timing intact.

Not verified: that it looks better. Human review accepted character 2x, custom 1.5x, game 2x and a 24-second 2x as "normal/acceptable", but the author's summary is that different enhancement methods give you different skin and texture character and none is universally best. Same with 4x - if the environment report says the candidate weights are missing for that scale on your model, the node will not fake it, and there's no independent community testing of this node to fall back on.

One trap worth stating: if your render is an H3 clip with generated dialogue, do an A/B before you trust the audio integrity checks to mean the mouths still line up. The checks prove the samples survived, not that the picture did the right thing. And bear in mind the special low-resolution safety pass: cropping, HUD text and heavy upscales are where you should look for texture smearing.

Install

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

Restart fully. Then install and licence Topaz Video yourself - the pack contains no program, model weights or licence files, downloads nothing, and doesn't accept any agreement on your behalf. Wire MiniMaxH3TopazEnvironmentEXPT8 → this node → a save/preview node.