HiAR proposes hierarchical denoising for autoregressive video diffusion models, a paradigm shift from conventional block-first to step-first denoising order. By conditioning each block on context at a matched noise level, HiAR maximally attenuates error propagation while preserving temporal causality, achieving state-of-the-art long video generation (20s+) with significantly reduced quality drift.
Create a conda environment and install dependencies:
conda create -n hiar python=3.10 -y
conda activate hiar
pip install -r requirements.txt
pip install flash-attn --no-build-isolation
# Download Wan2.1 base model
huggingface-cli download Wan-AI/Wan2.1-T2V-1.3B --local-dir-use-symlinks False --local-dir wan_models/Wan2.1-T2V-1.3B
# Download HiAR checkpoint
mkdir -p ckpts
wget -O ckpts/hiar.pt https://huggingface.co/jackyhate/HiAR/resolve/main/hiar.ptGenerate short videos (~5 seconds, 21 latent frames):
python inference.py \
--config_path configs/hiar.yaml \
--checkpoint_path ckpts/hiar.pt \
--data_path data/prompts.txt \
--output_folder outputs/ \
--num_output_frames 21 \
--use_ema \
--inference_method timestep_firstGenerate long videos (~20 seconds, 81 latent frames):
python inference.py \
--config_path configs/hiar.yaml \
--checkpoint_path ckpts/hiar.pt \
--data_path data/prompts.txt \
--output_folder outputs/ \
--num_output_frames 81 \
--use_ema \
--inference_method timestep_firstHiAR's hierarchical denoising structure naturally admits pipelined parallel inference: each GPU handles one denoising stage, and blocks flow through a diagonal pipeline schedule.
# Requires exactly N GPUs for N denoising steps (default: 4)
torchrun --nproc_per_node=4 \
scripts/pipeline_parallel_inference.py \
--config_path configs/hiar.yaml \
--checkpoint_path ckpts/hiar.pt \
--prompt "A cat sitting on a windowsill watching the rain fall outside" \
--output_path outputs/pipeline_output.mp4 \
--num_output_frames 81Our training algorithm is data-free (no video data is needed), requiring only text prompts and pre-generated ODE trajectory pairs from the teacher model for forward-KL regularization.
# Download Wan2.1 teacher model (14B, for real score computation)
huggingface-cli download Wan-AI/Wan2.1-T2V-14B --local-dir-use-symlinks False --local-dir wan_models/Wan2.1-T2V-14B
# Download ODE initialization checkpoint from the official Self-Forcing release
mkdir -p ckpts
wget -O ckpts/init.pt https://huggingface.co/gdhe17/Self-Forcing/resolve/main/checkpoints/ode_init.ptWe use the same VidProM-filtered prompt set as Self-Forcing. Download the prompt file from the Self-Forcing release:
mkdir -p data
wget -O data/vidprom_filtered_extended.txt https://huggingface.co/gdhe17/Self-Forcing/resolve/main/prompts/vidprom_filtered_extended.txtPre-generate ODE trajectory pairs from the teacher model. These serve as targets for the forward-KL regularizer that prevents low-motion shortcuts:
# Single-node, 8 GPUs (generates ~16K pairs by default)
bash scripts/run_generate_ode_pairs.sh
# Multi-node example
NNODES=4 NPROC_PER_NODE=8 NODE_RANK=0 MASTER_ADDR=<ip> \
bash scripts/run_generate_ode_pairs.shThe script supports resuming from interruptions with --resume. Adjust NUM_PROMPTS to control how many pairs to generate (10K-50K recommended).
torchrun --nnodes=8 --nproc_per_node=8 --rdzv_id=5235 \
--rdzv_backend=c10d \
--rdzv_endpoint $MASTER_ADDR \
train.py \
--config_path configs/hiar.yaml \
--logdir logs/hiar \
--disable-wandbInstead of fully denoising each video block before moving to the next (block-first), HiAR reverses the loop order to step-first: at each denoising step, all blocks are processed causally, with each block conditioned on previous blocks at the same noise level. This matched-noise context maximally attenuates inter-block error propagation.
To prevent the low-motion shortcut inherent to reverse-KL (DMD) distillation, a forward-KL regularizer is applied in bidirectional attention mode at the first denoising step, preserving motion diversity.
This codebase is built on Self-Forcing, CausVid, and Wan2.1.
@misc{zou2026hiarefficientautoregressivelong,
title={HiAR: Efficient Autoregressive Long Video Generation via Hierarchical Denoising},
author={Kai Zou and Dian Zheng and Hongbo Liu and Tiankai Hang and Bin Liu and Nenghai Yu},
year={2026},
eprint={2603.08703},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2603.08703},
}