H3 Progressive · HIGH Sampler Only (T8 EXP)

This is the one node in the progressive family that actually produces something you can watch. It runs the tail of the schedule at full canvas and hands you back a completed joint AV latent, ready for the pack's decode path.

"Only" in the name is doing real work. It runs HIGH. It does not run LOW, it does not run the learned upscale, it does not silently retry, and it will not reach back into a cache to find you a result.

What it does

Inputs: high_restart (from HIGH Handoff Only), the stage's model and sampler, its positive and negative conditioning, a seed, a cfg, and a reserve_vram_mib budget. Outputs: av_latent, report_json, and high_result.

Two implementation details are worth knowing because they explain behaviour you'd otherwise read as a bug:

• It executes the tail using explicit zero-noise reconstruction with the original clean source and masks. In plain terms: the starting state is rebuilt from the clean anchor rather than carried forward as a noisy tensor, which is what keeps HIGH anchored to the same trajectory the plan defined.
• seed here is the sampler seed, and under the original policy it's the same as LOW's - not the HIGH video-noise seed. Those are two different numbers that both matter, and the handoff node is where the noise one lives. If you're debugging a seed mismatch, check which one you're looking at.

cfg defaults to 1.0, consistent with the native Euler / CFG-1 route the plan is built around. reserve_vram_mib (default 1024) is a boundary check before the stage runs, in the same "seatbelt, not a guarantee" sense as on the LOW side.

Wiring

high_restart ─────────────→ HIGH Sampler Only ─→ av_latent ─→ decode
model, sampler ────────────↗                   └→ high_result ─→ Save / Effects Audit
positive, negative ────────↗

Both outputs matter. av_latent is the video-side payload; high_result is the typed completion evidence - path and receipt - that the save and audit nodes consume. Wire high_result into Save Completed HIGH AV if you plan to reload this run later, and into HIGH Effects Execution Audit if you applied any effects.

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

Then fully exit ComfyUI and start it again before reloading the page - the pack is explicit that new nodes need a Python reload, and "I refreshed and nothing changed" is nearly always an unrestarted process.

Nothing gets pip-installed. The requirements.txt is comments; the author keeps optional EXP dependencies out of it on purpose so an install can never replace ComfyUI's Torch/CUDA stack. Your side of the deal is a recent ComfyUI with native H3 support plus the H3 weights, Qwen text encoder and both VAEs in their standard folders.

Where people get burned

Two MODELs, one pipeline. HIGH can use its own model and LoRA chain - that's a feature - but architecture, latent format, video/audio shifts and audio scale have to match LOW. Weights may differ; shape compatibility doesn't guarantee your character survives the handoff.

Forgetting that this route samples audio jointly. Unlike the pack's "lock the first pass's audio" two-pass workflows, progressive keeps audio in the loop on both sides of the boundary. So a HIGH-stage prompt change or effect can move the sound, and any lip-sync or music review has to happen on the finished clip, not on the LOW boundary.

Reading the progress bar as work. The pack's own reporting distinguishes callback counts, model call counts and actual network forwards precisely because the progress bar doesn't tell you which ran. If your run felt suspiciously fast, check the report before celebrating.

Finally, the EXP framing, which the author is unusually candid about: verified speed and quality claims are limited to specific fixed clips on specific hardware - roughly a third off end-to-end time in the author's own A/B on an RTX 4060 Ti 16 GB, with the caveat that it's a cold-start single-sample result and not a general promise. Your call whether that's worth a lower-resolution first pass for your material. For iterating composition quickly, it usually is.