MiniMax H3 Acc LoRA: Official 8-Step PDD Distillation in ComfyUI

MiniMax-H3-Acc-LoRAs (Hugging Face | VideoX-Fun | arXiv) is an official release from Alibaba PAI applying Parallel Decoding Distillation (PDD) to MiniMax H3. Two acceleration adapters render joint video plus synchronized audio in 8 or 4 sampler steps with no classifier-free guidance, and a dedicated ComfyUI node pack brings the release to native ComfyUI workflows.

What it is

The two -Acc-8Step files are not ordinary LoRAs. Alongside a rank-64 trunk LoRA (network_alpha = 64, BF16), each checkpoint carries a Parallel Decoding Distillation head bank: 32 per-interval copies of the final-layer video and audio projections that are fused into a single mean-block-velocity head at each sampler step. A plain LoRA loader cannot read these files, and dropping the head bank silently loses the distillation.

8 steps, no CFG — generation runs at guidance_scale = 1.0 with a single forward pass per step; guidance is distilled into the adapter.
Audio + video in one pass — the distillation covers the joint audio stream, not just the visual frames.
Two trunks — one adapter targets the FL2VA (text/first-last-frame conditioned) checkpoint, the other the Ref2VA (reference conditioned) checkpoint.
Official release — published by Alibaba PAI on August 26, 2026, following the PDD method of Shaul et al. (arXiv 2607.26004).

Weights

File	Size	Target base	Link
`MiniMax-H3-FL2VA-Acc-8Step.safetensors`	1.4 GB	MiniMax-H3 (FL2VA)	HF
`MiniMax-H3-Ref2VA-Acc-8Step.safetensors`	1.4 GB	MiniMax-H3 (Ref2VA)	HF

Each file weighs 1.4 GB in BF16, rank 64, and is applied at strength 1.0 over the base H3 transformer. Pair the FL2VA adapter with an FL2VA UNET and the Ref2VA adapter with a Ref2VA UNET (bf16 originals or int8 convrot builds both work).

Results

Official demo samples generated with the Acc-8Step adapter at a LoRA weight of 1.0:

FL2VA

Text-to-video with the FL2VA Acc-8Step adapter

Ref2VA

Reference-to-video with the Ref2VA Acc-8Step adapter

The model card also ships side-by-side comparisons against the un-distilled baseline and the community Minimax-h3-Turbo 4-step LoRAs for both trunks.

ComfyUI support

ComfyUI-MiniMax-H3-PDD-Acc by Jalen-Brunson adds native ComfyUI nodes for the release. The MiniMaxH3PDDAccApply node applies the trunk LoRA and installs the PDD head bank on the final layer, armed per step by sigma, so looping and chunked samplers stay in sync. The companion scheduler node emits the trained sigma grid for partial-denoise workflows.

The trained recipe is strict:

Sampler — Euler; each step consumes one mean block velocity, so multi-stage samplers evaluate off-grid.
Steps — 8 (default, the trained block size), 4 (officially sanctioned regrouping), or 6 (non-uniform 8,8,4,4,4,4 partition). Other step counts are rejected by the node instead of silently degrading.
Guidance — CFG 1.0 with BasicGuider.
Sigma shift — exactly 12.0 / 3.0.
Pruned checkpoints — on pruned (curve-table) H3 UNETs the 50 dense-adaLN LoRA modules are automatically rebased onto the model's curve basis.

Remove other distill LoRAs (lightx2v turbo etc.) when using these adapters — distillations do not stack. Character LoRAs stack normally.

Two formats are supported: the original Alibaba files (converted in memory by the node) or the pre-converted ComfyUI-key redistribution by aptech0081/MiniMax-H3-Acc-LoRAs-ComfyUI (1.7 GB per file).

Availability

Outside ComfyUI, the adapters run through the VideoX-Fun inference pipeline. Set model_path and pdd_lora_path in the predict_t2v.py (FL2VA) or predict_ref2v.py (Ref2VA) example scripts from VideoX-Fun, which use Diffusers' MiniMax-H3 ModularPipeline and require diffusers >= 0.40.0. Each example loads the checkpoint with apply_pdd_lora and derives the required number of inference steps from its configuration.