H3 Chunked v5 · Global Learned Lift (T8 EXP)

If you remember one structural difference between the old chunked route and the v5 one, make it this: v5 upscales the entire clip in a single learned 3D lift, and only then cuts the timeline into refine windows. The older v1–v4 route sliced first and lifted per segment.

That "lift once, globally" decision is what makes the windows cheap to reason about later - every window refines the same continuous high-resolution AV latent instead of its own separately-upscaled island. It's also why v5's nodes reject the older per-segment source path outright.

H3 Chunked v5 · Global Learned Lift is that single upscale. Two inputs: partial4_denoised_output (the verified low-res half-finished AV latent, straight from Verify Native Partial4 Source) and plan (the chunked two-pass plan that carries your target geometry). Three outputs: lifted_full_av, the typed lift_receipt, and report_json.

Use your normal learned 3D upscaler weights - the node runs "the original learned 3D upscaler", so it's the same network you'd use in the all-in-one chunked sampler. Nothing is downloaded for you.

Why the receipt exists

lift_receipt is a typed T8_CHUNKED_V5_GLOBAL_LIFT, and Prepare Full AV Noise requires it as an input. That's not decoration. The v5 contract is that exactly one global lift happened, over exactly this source and this plan, and the prepare node wants to know it before it draws the global noise field.

The node's own description volunteers what it does not do: it doesn't sample PASS2 and it doesn't produce a portable cache receipt. So this is a step in a chain, not a checkpoint. If you want a resumable artifact, that's MiniMaxH3ChunkedV5WindowSaveEXPT8 after a window has actually been sampled.

What the numbers look like in practice

From the pack's own standard-4+4 documentation, a real 0.401 MP eight-second run at 192 frames, 24 fps, uses a 136-frame temporal window with 34-frame overlap, giving two windows ([0,136) and [102,192)). That's a shared four-step low pass plus four steps per window: twelve model forwards total, not eight. "4+4" is the recipe for each region, not the whole-clip total. Memory-wise that run recorded roughly 4.3 GiB free VRAM at its lowest sample point, and the author is explicit that this is an observation, not a promise - no general 16 GB guarantee exists.

The point of the temporal window is to cut the activation peak of the high-resolution transformer. It does not shrink the model weights, the lifted latent or the noise field, all of which still sit in memory.

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

Fully restart ComfyUI, then refresh the tab. Recent Core with native H3 support is required; the Registry listing and GitHub releases ship independently, so clone from GitHub if Manager lags. The pack adds no pip dependencies by design - requirements.txt is empty so an install can never replace your Torch/CUDA stack. The learned 3D upscaler weights, TAEH3 preview models and friends live on the author's HuggingFace under t8star; main model, Qwen text encoder and video/audio VAEs follow the README's model table.

Where it goes wrong

Wrong source. Feed it a first-pass latent from the old v1 route and it has no verified partial4 identity to bind. It will complain rather than guess.

Geometry drift. plan decides the lift geometry. If you changed the target size in the plan after building the graph, the downstream prepare node's re-checks won't line up.

Doing it twice. One lift, one plan, one receipt. Two lift nodes feeding one prepare node is not a supported shape.

Duplicate installs. If panels don't match the example workflow, check custom_nodes for a stale second copy of the pack. Rename it with a .disabled suffix (a leading underscore won't do it), restart, and confirm the python_module in /object_info.

Seam behaviour, lip-sync and voice still need eyes and ears on the finished clip. The pack's own README lists a slight seam colour shift as a known limitation of dual-pass stitching.