HiAR changes autoregressive video diffusion from block-first denoising to a step-first hierarchical schedule. Each block conditions on earlier blocks at the same noise level, which reduces long-horizon drift while preserving temporal continuity. This dependency pattern also facilitates pipelined parallel inference, while a dedicated forward-KL regulariser maintains motion diversity throughout self-rollout distillation—an essential requirement for training within this denoising paradigm.

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.pt

Generate 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_first

Generate 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_first

HiAR'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 81

Our 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.pt

We 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.txt

Pre-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.sh

The 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-wandb

In essence, this method allows autoregressive video generation to mimic a bidirectional attention video denoising paradigm. For instance, the high-noise denoising stages only require coarse-grained context information. This design maximally reduces error accumulation while theoretically retaining sufficient information to maintain continuity. By scaling the training budget under the constraint of the Forward KL loss, we can achieve near-zero degradation in most scenarios, even enabling infinite generation (e.g., over 200 minutes). However, in some dynamic scenes, inter-frame jumping may still occur. We believe this is not an inherent limitation of the hierarchical denoising paradigm itself, but rather an issue of insufficient capacity in the 1.3B base model, as this denoising paradigm is considerably more challenging. We plan to further validate this paradigm on more powerful base models in the future.

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}, 
}