H3 Continuation · LOW Sampler Only (T8 EXP)

The two-pass trick that makes H3 long video affordable is simple to say and fiddly to own: draw the first steps at half the canvas, upscale the latent with a learned model, finish at full resolution. Progressive sampling numbers from the pack's own testing on a 4060 Ti put the win at roughly 35–38% end-to-end versus a straight 8-step pass - real, but not the 4x people assume from "half the width and half the height".

This node is the first half, and only the first half. If you want the illusion of a black box, this is the wrong pack.

What it does

It runs the native Euler low-resolution pass on this segment, using the original accepted clip's frames as motion guides, at CFG1 the way H3 wants it. Then it stops, and instead of handing you a video it hands you a typed boundary: clean video in model space plus the evolving audio state. Not a decoded image, not a finished latent.

That boundary is deliberately not decodable. Anything you could hand a VAE decoder at this point would be a lie - the resolution is wrong, the schedule is half-run, and the audio is mid-flight.

The other half of the deal is independence. The node takes its own model, its own sampler, its own noise and its own conditions. You can run a different LoRA stack or a different attention backend on the low pass than you do on the high pass, which is genuinely useful - cheap steps tolerate more abuse. What you cannot do is change the architecture or the AV coordinate layout between the two halves; the HIGH stage has to continue the same maths.

Inputs that matter

• prepared_phase - from the "ONE Phase Conditions" node with its phase set to low. Feed it a HIGH phase and it refuses.
• plan - the object from the Plan node. It carries the frame count, canvas and the low sigma half.
• noise - both the seed and the actual noise tensor. This seed is the one the HANDOFF stage needs, because the pack's original noise policy derives the HIGH video noise from it as seed+1.
• model, sampler, positive, negative - the usual, plus the plan's low-half schedule.
• reserve_vram_mib - default 1024. It is a headroom check at stage boundaries, not a memory limit for your card, and the docs are clear it does not promise freedom from OOM.

Outputs are low_boundary, the typed object that goes to the progressive lift input and then to the HIGH handoff, and report_json, which records the model calls, the clock state and the receipt hash.

Where it sits

Parent → context preparation → Plan → conditions (low) → LowStage → lift → learned 3D latent upscaler → HIGH handoff → HIGH stage → decode.

Note the audio behaviour, because it is the thing that surprises people coming from the older two-model long-video node: the audio is not frozen here. The low pass seeds the audio clock and the high pass keeps evolving it. That is why music and lip-sync can change when you switch to this route, and why you have to listen to the finished segment rather than trusting the previous workflow's result.

Installing it

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

Full restart afterwards - module import, not page refresh. Or search "MiniMax H3 Audio T8" in ComfyUI Manager. The pack has no pip dependencies by design; the base nodes use ComfyUI's own torch/numpy/Pillow, and optional EXP features load their own requirements only when invoked. You still need a recent Core with native H3 support, plus the H3 checkpoint, Qwen text encoder, video VAE and audio VAE.

Things that will bite you

low_boundary is not video. Wiring it into a decode or a save node produces either an error or something ugly, depending on the node's patience.

If the LOW pass and the HIGH pass use the same seed by accident when the original policy wants the low seed plus one for the high video noise, the seam shows up as a colour or grain shift around the boundary - a known limitation the pack documents rather than hides. Check your seeds before you blame the upscaler.

And a boring one: the whole route is marked experimental in this pack's own docs. Small random-model and CPU checks are not quality validation, and the author says the same in every release note. Run one segment, watch the seam and listen to the audio, before you queue a 32-second job.