pi-H3
video with audio in 2–4 steps without VAE decoder
2 min readupdated
A set of acceleration adapters from researchers at Stanford University and Luma AI for the MiniMax-H3 video model. They enable generating video with synchronized audio in just 2 or 4 diffusion steps. The standout feature of DWD (Decoding Without a Decoder) versions is that the model predicts final frame pixels directly in the final step, completely bypassing the heavy VAE decoder. These adapters are useful for anyone wanting to run MiniMax-H3 significantly faster via diffusers or ComfyUI.
Key features
- 2 or 4-step generation — thanks to pi-Flow v2 distillation, the process requires minimal iterations instead of dozens of passes from the base model.
- Direct pixel output (DWD) — adapters labeled
dwdgenerate frames directly, so VAE is only used to encode keyframes. - Synchronized audio — the model outputs a video track together with audio right away (24 fps frequency).
- Three operation modes — text-to-video generation is supported, as well as generation from a first (
image) or last (last_image) frame. - Ready-to-use versions for ComfyUI — quantized INT8 ConvRot variants are available for DWD adapters.
What you can do
- Generate 10-second clips at 1344×768 resolution with sound accompaniment based on a text prompt.
- Bring static images to life or build smooth transitions to a specified final frame.
- Assemble a fast local pipeline in ComfyUI via
ComfyUI-piFlowusing lightweight INT8 files.
Example
from lakonlab.pipelines.pipeline_minimax_h3 import MiniMaxH3DWDBlocks
from lakonlab.pipelines.utils import load_lakonlab_adapter
pipe = MiniMaxH3DWDBlocks().init_pipeline('multimodalart/MiniMax-H3-Pruned')
load_lakonlab_adapter(pipe, 'Lakonik/pi-H3', subfolder='pi_h3_4step_dwd_adapter')
out = pipe(prompt='A red fox trotting through a snowy pine forest',
num_inference_steps=4, output=['dwd_pixels', 'audio', 'sampling_rate'])Read more
DetailsVariants:
pi_h3_2step_dwd_adapter, pi_h3_4step_dwd_adapter, as well as classic versions for the VAE decoder.Optimizations: the diffusers example uses FP8 W8A8 quantization and Sol sparse attention.
Limitations:
int8_convrot.safetensors files only work in ComfyUI; MiniMax H3 Community License.













Comments