H3 PDD · Separate Absolute LOW4 / HIGH4 (T8 EXP)

PDD - Parallel Decoding Distillation - is Alibaba PAI's 8-step acceleration for MiniMax H3: 32 time intervals collapsed into 8 model forwards, each forward consuming the weighted mean of four consecutive absolute output heads. It's the same family of trick as Lightning or DMD, just weirder, because the adapter isn't only a LoRA. It ships 32 video heads and 32 audio heads alongside the backbone adapters, which is why a plain Load LoRA node quietly does nothing useful with it.

This node exists because people want to run those 8 forwards across two resolutions: heads 0–3 at a small canvas, a learned latent upscale, then heads 4–7 at full size. Separate Absolute LOW4 / HIGH4 sets up exactly one of those halves. It runs no sampler at all.

What it actually does

Connect the MODEL that already came out of the existing PDD 8-Step Setup node, plus that node's original full nine-point SIGMAS. You get back a new MODEL branch, a SAMPLER, the slice of SIGMAS for this half, and a typed stage context.

The mechanism is deliberately boring: the node reads the PDD attachment receipt off the incoming model, checks it's a genuine t8_minimax_h3_pdd_8step_setup_v2 receipt, validates the shape of the head banks (and their finiteness), then rebuilds the native sampler geometry for this stage's actual AV geometry while keeping the nine-point table intact. Learned upscaling and joint-audio reconciliation stay outside, on purpose - that's the seam you're choosing, not something the node should hide.

If you feed it a plain H3 model, it fails with Connect MODEL from the existing PDD 8-Step Setup. That's the node working correctly.

Inputs that matter

• model - from the PDD 8-Step Setup. Each half may get its own LoRA chain and attention backend, as long as the architecture and AV clock actually match.
• full_sigmas - the original full nine-point table. Not a pre-split one; this node does the splitting.
• av_latent - the empty (or initialized) joint audio/video latent for this stage.
• stage - pdd_low_0_4 for heads 0–3, pdd_high_4_8 for heads 4–7.

Outputs: model, sampler, sigmas, stage_context (feed the Stage EAV / audit nodes), and report_json. Model, sampler and sigmas go into a SamplerCustomAdvanced. Nothing here executes on its own - you still own the sampler node.

The 4+4 split is not "8 steps, twice." Total transformer calls across both halves is still 8, one per output-head block. If you configured 8 steps and split at 4, LOW takes the first four evaluations and HIGH takes the last four.

Install

ComfyUI 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 restart ComfyUI and refresh the browser page. The pack's requirements.txt installs nothing on purpose - it won't touch your Torch/CUDA stack. Optional EXP features check their own dependencies only when you actually use them.

For PDD specifically you need the converted adapter in ComfyUI/models/loras (MiniMax-H3-FL2VA-Acc-8Step_comfyui_pdd.safetensors, or the Ref2VA variant) plus the corresponding full, non-pruned base model. The stock H3 video/audio VAE and Qwen text encoder still have to be in their usual folders - PDD doesn't replace any of that.

Where people get burned

The cold-HIGH trap. The split graphs come in two flavours: Full_Stages, which runs and saves both halves, and Cold_HIGH, which reads a saved LOW by artifact_path + artifact_sha256. Those fields in the shipped JSONs are placeholders. Fill them from your own run, and re-run LOW whenever you change the LOW model, LoRA, prompt, first frame or canvas - a matching SHA proves the bytes, not that this morning's settings match yesterday's.

Also: the registry listing and the GitHub release move independently. If Manager doesn't show the version whose docs you're reading, install from GitHub.

None of this is a general performance promise. The author's own docs are unusually blunt about the verification boundary - fixed test clips, specific hardware - and 16 GB cards are not guaranteed. Treat that honestly; it's why the pack is labelled EXP.