H3 PDD Acceleration (8-Step)

MiniMax H3 is a 33B omni-modal video model that makes picture and stereo audio in one pass, and it denoises like a big video model - slowly. The PDD acceleration files from Alibaba PAI are the fix, and this node is how you actually run them in ComfyUI. It is not a LoRA loader, even though the files sit in your models/loras folder and end in .safetensors. Treat it as one and you'll spend an evening confused.

PDD, in one paragraph

PDD is parallel decoding distillation: instead of training a student to imitate one teacher step at a time, you train it to predict several steps at once. NVIDIA published the method as FastGen-PDD and demoed it on LTX 2.3; Alibaba's MiniMax-H3-Acc-LoRAs are H3 distilled the same way. Same family as the CausVid and LightX2V step-distill LoRAs the Wan crowd stacks - except those are plain adapters, and these are not.

Why the files break the normal loader

Two things are in each file: rank-64 backbone adapters, and a full bank of schedule-dependent output heads. ComfyUI's LoRA loader has no concept of the second part, because nothing else ships one. So the node does the work itself:

• maps 362 Diffusers-format rank-64 adapters onto native ComfyUI H3 keys;
• folds Q/K/V adapters into the right slices of ComfyUI's fused QKV weights;
• fuses 32 displacement heads - four adjacent intervals integrated into one Euler velocity - into eight schedule-weighted video and audio heads, in fp32, once, on CPU;
• keeps exactly one fused head on the GPU at a time (a couple of MB) and picks it from the actual sampler sigma, not a step counter;
• emits the matching eight-NFE sigma schedule as its second output.

That last point is the reason the node has two outputs instead of one.

Wiring it

model is a native MiniMax-H3 MODEL, straight off your diffusion-model loader. pdd_lora_name is where you point at the PDD file - pick MiniMax-H3-FL2VA-Acc-8Step.safetensors for an FL2VA or hybrid base, MiniMax-H3-Ref2VA-Acc-8Step.safetensors for a full Ref2VA base. It's a dropdown of everything in models/loras, so a wrong pick shows up as garbage downstream, not as an error.

Load Diffusion Model -> H3 PDD Acceleration (8-Step) -> SamplerCustom MODEL
                                      SIGMAS (8 NFE) -> SamplerCustom SIGMAS
KSamplerSelect: euler                              CFG: 1.0

strict_schedule stays on. It makes the node reject any sampler schedule that isn't the exact eight distilled blocks, which sounds pedantic and is actually the difference between the intended model and an approximation of it. Leave allow_experimental_shifts off too: Alibaba distilled these heads at video/audio shifts 12/3, and MiniMaxH3SigmaShift - if you use it - has to say 12/3 and sit before this node. If you do use the shift node elsewhere and chain it after, you'll get an error naming both shift pairs; that's the node telling you which way round they go, not a bug.

hybrid_mode and hybrid_ref2va_pdd_lora are for smhfacct hybrid checkpoints (b15/b20/b25/b30-49). Hybrid keeps FL2VA adapters and heads everywhere except the late adaln_proj adapters, which come from Ref2VA - structurally faithful, never jointly distilled by Alibaba, so it's experimental. Manual plus manual_hybrid_start_block is the same thing for a boundary that has no name.

Install

ComfyUI Manager, search "MiniMax H3 PDD Acceleration". Or:

cd ComfyUI/custom_nodes && git clone https://github.com/lukas-9936/ComfyUI-MiniMax-H3-PDD

Then drop the two Alibaba PDD .safetensors into ComfyUI/models/loras and restart. There's no requirements.txt - no extra Python packages, just torch and ComfyUI itself. The two h3_silu_temb_grid_*.safetensors files (about 11MB each) are bundled in the repo and found automatically, so don't go hunting for them; they're support assets, not LoRAs. An older generic h3_silu_temb_grid.safetensors is obsolete once both named grids exist.

Where people get burned

The errors are explicit, and they almost all mean you wired it wrong:

• "requires its exact eight-step SIGMAS output." You put something between this node and SamplerCustom, or set the sampler's steps to anything other than 8. Wire the SIGMAS output directly.
• "received sigma … which is not one of its eight distilled sampling points." Same cause from the other side - a different step count, or a scheduler node re-deriving the schedule.
• "H3 PDD Acceleration is already applied to this MODEL." Two PDD nodes on one chain. One is enough.
• Curve/pruned checkpoints. The node rebases its 50 dense AdaLN adapters internally; you don't need the AdaLN Fix node for this output. If it can't find a trustworthy dense time curve it stops and tells you to drop a non-pruned H3 checkpoint in models/diffusion_models or matching grids in models/h3_adaln. Dense checkpoints ignore grids entirely.
• The node doesn't show up. It registers through comfy_api.latest, so it needs a current ComfyUI; on an old build startup logs [H3PDD] failed to load. And don't leave an older ComfyUI-H3-PDD-Acceleration nested inside another pack - both copies register the same node ID.

One expectation to set: combining PDD with SLA or an SLA Turbo LoRA is off-distribution, and Spectrum isn't supported - PDD already defines its own eight evaluations. The combo worth testing is KJNodes' low-VRAM H3 attention plus chunked feedforward.