Accelerated Video & Audio Generation in 8 Steps
Parallel Decoding Distillation (PDD) enables the massive 33B parameter MiniMax-H3 model to synthesize high-resolution 768p/1080p video with synchronized stereo soundtracks in only 8 inference steps with minimal degradation.
Side-by-Side Quality & Speed Comparison
Compare Baseline MiniMax-H3, 4-Step Turbo, and Alibaba's 8-Step Acc LoRA on official test cases
Video Synthesis Parameters
Fixed to 8 steps by PDD config
Video Output Player
Native 24 FPS · Stereo Synchronized AudioSelect prompt and parameters then click Synthesize.
How Parallel Decoding Distillation (PDD) Works
Standard video diffusion models require 50+ sequential Euler steps because each step relies on the output of the previous step. Parallel Decoding Distillation (Shaul et al., 2026) circumvents this bottleneck by repeating the output projection heads across an \(N\)-step time grid (\(N=32\)), grouped into blocks (\(L=4\)).
Multi-Head Output Fusing
Instead of a single linear layer for final latent predictions, PDD equips the transformer with MiniMaxH3ParallelHead for both video (proj_out) and audio (audio_proj_out). Each generation step dynamically evaluates a block of intervals simultaneously:
Shifted Time Grid & Sampling Plan
MiniMax-H3 uses asymmetric noise schedule shifts for video (\(\text{shift}=12.0\)) and audio (\(\text{shift}=3.0\)). The PDD sampling plan computes the mean velocity over each block span so a single Euler update covers the entire 4-step interval:
"Parallel Decoding Distillation for Fast Image and Video Generation", arXiv:2607.26004
Programmatic Quickstart
import torch
from diffusers import ComponentsManager, ModularPipeline
from minimax_h3_pdd import apply_pdd_lora
# 1. Initialize MiniMax-H3 Modular Pipeline
model_path = "MiniMaxAI/MiniMax-H3"
manager = ComponentsManager()
manager.enable_auto_cpu_offload(device="cuda", memory_reserve_margin="12GB")
pipeline = ModularPipeline.from_pretrained(model_path, workflow="t2va", components_manager=manager)
pipeline.load_components(dtype=torch.bfloat16, pretrained_model_name_or_path=model_path)
# 2. Attach Alibaba PAI 8-Step Acc LoRA
nfe = apply_pdd_lora(
pipeline.transformer,
checkpoint="alibaba-pai/MiniMax-H3-Acc-LoRAs",
video_shift=pipeline.scheduler.shift,
audio_shift=pipeline.audio_scheduler.shift
)
# 3. Generate in 8 Steps!
result = pipeline(
prompt="A majestic eagle soaring over snow-capped mountains at sunset.",
height=704,
width=1280,
num_frames=124,
num_inference_steps=nfe + 1, # 8 + 1 = 9
generator=torch.Generator().manual_seed(42),
output=["videos", "audio", "sampling_rate"]
)
Hugging Face offers free community GPU grants for open-source AI projects. Since this Space already includes complete app.py and minimax_h3_pdd.py ZeroGPU code, you can apply in 1 click from your Space Community tab!