MiniMax H3 PDD 8-Step Setup (T8 Advanced EXP)

MiniMax H3 is a 33B omni-modal model, and full-quality sampling is not quick. The PDD 8-Step Setup node is the T8 pack's answer for the people who want a real speedup without pretending distillation is free. It loads Alibaba PAI's converted MiniMax-H3 PDD acceleration adapter and, critically, does it in a way an ordinary Load LoRA node physically cannot.

Why it can't be a normal LoRA

PDD stands for the acceleration adapter from Alibaba PAI's MiniMax-H3 Acc-8Step family, and the converted *_comfyui_pdd.safetensors files are not your usual LoRA. Each file packs 258 backbone adapter triples plus dynamic video and audio output heads - 32 video heads and 32 audio heads that get fused into the runtime grid. A generic LoRA loader drops those dynamic heads and hands you a half-applied adapter. This node exists precisely so that doesn't happen.

The mechanism, from the source: it applies all 258 adapters as model-only bypass residuals - the INT8 base is left unmodified rather than re-quantized - patches the final layer with the eight pre-fused PDD heads, and returns a ready-made Euler/simple 8-step sampler with video/audio sigmas at the shift 12 / 3 the adapter was trained for. CFG 1, per the README. It also hard-rejects the stuff that will silently produce garbage: pruned AdaLN-curve bases, mixing FL2VA and Ref2VA adapters, stacked external LoRAs, and any non-official sigma grid.

The inputs that matter

• model - connect the matching native ComfyUI full, non-pruned H3 MODEL. The tooltip is emphatic: don't attach an existing LoRA, a pruned model, or another PDD node.
• av_latent - from the existing T8 Conditioning node; it configures ComfyUI's native joint audio/video sampler.
• pdd_lora_name - the converted file in models/loras, defaulting to the FL2VA variant.
• base_variant - FL2VA or Ref2VA. Metadata mismatches are reported but don't block, so you can shoot yourself in the foot quietly; pick the one matching your base.
• strength - the official setting is 1.0. Anything lower interpolates the residuals and the heads and is explicitly unvalidated. Leave it alone.

Outputs: model, sampler, sigmas, and report_json. The sampler and sigmas wire straight into a KSampler (or the dual-clock sampler), and report_json will tell you exactly how many adapters mapped - 258 is the number you want to see.

Installing it

The node comes with the pack, but the adapter is a separate download:

# from https://huggingface.co/t8star/MiniMax-H3-Acc-8Step-comfy
# put the FL2VA or Ref2VA *_comfyui_pdd.safetensors file into:
ComfyUI/models/loras

Then either clone the repo (git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8 inside custom_nodes) or install "MiniMax H3 Audio T8" via ComfyUI Manager, update ComfyUI itself first, and restart. The ready-to-import workflows live in examples/workflows/19-pdd-acceleration.

Where people get burned

Two mistakes dominate. First, treating the PDD file as a normal LoRA - loading it with Load LoRA silently drops the dynamic heads and you get a model that looks patched but behaves half-applied. Second, forgetting the base contract: pruned bases, the wrong variant, or an existing LoRA stack all get rejected for a reason, and the node's refusal is the friendlier outcome.

Also manage your expectations on memory. The author's own real 736x416x124 validation runs hit ~15.6GB peak VRAM and failed his own 512MiB residual-VRAM safety gate - FL2VA left 447MiB free, Ref2VA 510MiB. There is no "this makes 16GB cards safe" claim anywhere in the evidence, and per the distillation lore, 8 steps is 8 steps: faster, not strictly free. Start with the stock 8-step workflows, eyeball the first render before you trust it, and keep a full-sampling fallback around.