Extensions/MiniMax H3 PDD Acceleration
ComfyUI Extension

MiniMax H3 PDD Acceleration

Native ComfyUI loader for Alibaba PAI MiniMax-H3 PDD acceleration LoRAs

By lukas-9936·Created about a month ago·Updated about a month ago· 2
lukas-9936/ComfyUI-MiniMax-H3-PDD
Nodes—
On cloudLocal install
Stars2
Updatedabout a month ago
Readme

ComfyUI-MiniMax-H3-PDD

Standalone native ComfyUI integration for Alibaba PAI's MiniMax-H3 PDD acceleration LoRAs. It uses the original .safetensors files directly; no converted model file is required.

Installation

Copy or clone this folder as:

ComfyUI/custom_nodes/ComfyUI-MiniMax-H3-PDD

Put the acceleration files in ComfyUI/models/loras:

  • MiniMax-H3-FL2VA-Acc-8Step.safetensors
  • MiniMax-H3-Ref2VA-Acc-8Step.safetensors

For universal curve/pruned-model support, both explicitly named grid files are bundled in this custom-node folder:

  • h3_silu_temb_grid_fl2va.safetensors
  • h3_silu_temb_grid_ref2va.safetensors

The node discovers the bundled copies automatically, so a normal clone/copy of the complete repository needs no separate grid download. They are small support assets, not LoRAs. You may alternatively place them in ComfyUI/models/h3_adaln. An older generic h3_silu_temb_grid.safetensors is no longer needed once both named grids are installed.

Restart ComfyUI. The pack registers two nodes:

  • H3 PDD Acceleration (8-Step) loads the Alibaba PDD files and supplies their exact sigma schedule.
  • H3 AdaLN LoRA Fix (PDD Standalone) ports incompatible AdaLN patches already attached by an ordinary LoRA loader.

This custom node is independent of PlagueKind Nodes. Do not keep an older copy of ComfyUI-H3-PDD-Acceleration nested inside another custom-node pack, because both copies register the same node ID.

The standalone AdaLN node intentionally uses the unique node ID H3PDDStandaloneAdaLNFix, so it may coexist with PlagueKind's AdaLN node. Do not put both AdaLN-fix nodes on the same MODEL chain; choose either one.

Sampling

Wire both outputs from H3 PDD Acceleration (8-Step) to SamplerCustom:

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

Keep strict_schedule enabled. The supplied SIGMAS already cover the complete denoise-1.0 trajectory; do not place another scheduler between the PDD node and SamplerCustom.

The released PDD checkpoints were distilled at video/audio shifts 12/3. The node enforces those shifts by default. Recomputing fusion weights and sigmas for another shift does not retarget the learned backbone adapters or interval heads. allow_experimental_shifts is an explicit off-distribution escape hatch, not a quality-preserving conversion.

If the workflow uses MiniMaxH3SigmaShift, set it to 12/3 and place it before the PDD node. Otherwise native H3's default 12/3 settings are used.

Full and hybrid models

| Loaded model | Primary PDD | Hybrid mode | Secondary PDD | | --- | --- | --- | --- | | Full FL2VA | FL2VA | Off | None | | Full Ref2VA | Ref2VA | Off | None | | Hybrid b15-49 | FL2VA | b15-49 | Ref2VA | | Hybrid b20-49 | FL2VA | b20-49 | Ref2VA | | Hybrid b25-49 | FL2VA | b25-49 | Ref2VA | | Hybrid b30-49 | FL2VA | b30-49 | Ref2VA |

Hybrid mode mirrors the hybrid checkpoint: FL2VA adapters and output heads are used everywhere except adaln_proj adapters from the selected block through block 49, which come from Ref2VA. This mapping is structurally faithful but was not jointly distilled by Alibaba, so it remains experimental.

Curve/pruned H3 checkpoints

The node internally rebases all 50 dense PDD AdaLN adapters for curve-form H3 checkpoints. It does not register or require a separate AdaLN Fix node.

The separately exposed H3 AdaLN LoRA Fix (PDD Standalone) is for non-PDD LoRAs loaded through ComfyUI's ordinary LoRA loaders. Put it after those loaders. It uses the same self-contained porting implementation but is not required on the PDD node's own output.

The rebase needs a matching dense time curve. It discovers one from:

  • a matching non-pruned H3 checkpoint in models/diffusion_models; or
  • one or more h3_silu_temb_grid*.safetensors candidates in models/h3_adaln, models/loras, or another ComfyUI model search folder.

Use h3_silu_temb_grid_fl2va.safetensors for FL2VA-based curve models and the b15/b20/b25/b30 hybrids, and h3_silu_temb_grid_ref2va.safetensors for Ref2VA-based curve models. Both may be installed together: the node measures every candidate against the loaded checkpoint's own table and automatically uses the best safe fit. Dense models ignore all curve grids. If no candidate fits safely, the node stops rather than silently dropping the AdaLN adapters.

What the node does

  • maps 362 Diffusers-format rank-64 adapters to native ComfyUI H3 keys;
  • applies Q/K/V adapters to the correct slices of ComfyUI's fused QKV weights;
  • fuses 32 displacement heads into eight schedule-weighted video and audio heads in fp32;
  • transfers only the active head to the GPU;
  • selects it from the real sigma, not an invocation counter;
  • emits and validates the exact eight-NFE sigma schedule.

The files cannot be reduced to ordinary LoRAs for ComfyUI's standard LoRA loader: their schedule-dependent full output-head banks are essential to PDD.

Compatibility

  • KJNodes' H3 low-VRAM attention and chunk-feedforward execution patches are the most sensible combinations to test first.
  • PDD was distilled with dense H3 attention. Combining it with SLA or an SLA Turbo LoRA is off-distribution and needs visual/audio/reference/sync testing.
  • Spectrum changes which evaluations execute and is not currently supported; PDD already defines its own eight-evaluation trajectory.

No PDD model weights are included in this repository. The two small bundled curve grids are support assets derived from the matching MiniMax-H3 time curves. PDD assets and the reference implementation belong to Alibaba PAI and their respective authors. This independent ComfyUI integration is MIT-licensed.

Acknowledgements