Shadowgraph
About
Sign in

HKUST-SAIL/NoiseAR

Official implementation of "NoiseAR: AutoRegressing Initial Noise Prior for Diffusion Models"

18

stars

14

commits

Python

primary language

Jun 3, 2025

updated

README

NoiseAR: AutoRegressing Initial Noise Prior for Diffusion Models

Zeming Li$^{*}$ Xiangyue Liu$^{{*}}$, Xiangyu Zhang, Ping Tan, Heung-Yeung Shum

Getting Started

1. Download Documents for Setting Up the Environment, Our Pretrained Models, and Eval Datasets (Compulsory)

bash download.sh

2. Set Up Your Development Environment

conda create --name noise_ar python=3.10.8 
conda activate noise_ar
make install

3. Run Inference

For example, to run inference using SDXL with our trained model:

python try.py

4. Run Evaluation Metrics

For example, run eval using SDXL with our trained model on DrawBench Evaluation Dataset:

torchrun eval.py exp.suffix=eval_sdxl_DrawBench exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/DrawBench
Click to expand for all of our evaluation commands 
# SDXL on DrawBench, GenEval, PickaPic
torchrun --master_port 13340 eval.py exp.suffix=eval_sdxl_DrawBench exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/DrawBench
torchrun --master_port 13341 eval.py exp.suffix=eval_sdxl_GenEval exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/GenEval
torchrun --master_port 13342 eval.py exp.suffix=eval_sdxl_PickaPic exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/PickaPic

# DreamShaper on DrawBench, GenEval, PickaPic
torchrun --master_port 13343 eval.py exp.suffix=eval_dreamshaper_DrawBench exp.pipeline=DreamShaper exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/DrawBench exp.cfg=3.5
torchrun --master_port 13344 eval.py exp.suffix=eval_dreamshaper_GenEval exp.pipeline=DreamShaper  exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/GenEval exp.cfg=5.0
torchrun --master_port 13345 eval.py exp.suffix=eval_dreamshaper_PickaPic exp.pipeline=DreamShaper  exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/PickaPic exp.cfg=5.0

# DiT on DrawBench, GenEval, PickaPic
torchrun --master_port 13346 eval.py exp.suffix=eval_dit_DrawBench exp.pipeline=DiT exp.pretrained_path=pretrained_models/dit/model.pth exp.val_data_path=data/DrawBench 
torchrun --master_port 13347 eval.py exp.suffix=eval_dit_GenEval exp.pipeline=DiT  exp.pretrained_path=pretrained_models/dit/model.pth exp.val_data_path=data/GenEval
torchrun --master_port 13348 eval.py exp.suffix=eval_dit_PickaPic exp.pipeline=DiT  exp.pretrained_path=pretrained_models/dit/model.pth exp.val_data_path=data/PickaPic 

# DPO of SDXL, DreamShaper, DiT on DrawBench
torchrun --master_port 13349 eval.py exp.suffix=eval_sdxl_DrawBench_DPO exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper_dpo/model.pth exp.val_data_path=data/DrawBench
torchrun --master_port 13350 eval.py exp.suffix=eval_dreamshaper_DrawBench_DPO exp.pipeline=DreamShaper exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper_dpo/model.pth exp.val_data_path=data/DrawBench
torchrun --master_port 13351 eval.py exp.suffix=eval_dit_DrawBench_DPO exp.pipeline=DiT exp.pretrained_path=pretrained_models/dit_dpo/model.pth exp.val_data_path=data/DrawBench exp.cfg=5.0

🚧 Todo

  • Release the training code & data.
  • Release the training code & data for DPO.

📍 Citation

If you find this project useful for your research, please cite:

@misc{li2025noisearautoregressinginitialnoise,
      title={NoiseAR: AutoRegressing Initial Noise Prior for Diffusion Models}, 
      author={Zeming Li and Xiangyue Liu and Xiangyu Zhang and Ping Tan and Heung-Yeung Shum},
      year={2025},
      eprint={2506.01337},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2506.01337}, 
}

Contributors

Lxiangyue

14 commits

Similar repos

xie-lab-ml/Golden-Noise-for-Diffusion-Models

[ICCV2025] The code of our work "Golden Noise for Diffusion Models: A Learning Framework".

197

Klayand/Golden-Noise-for-Diffusion-Models

Inference code of "Golden Noise for Diffusion Models: A Learning Framework".

4

Kwai-Kolors/LPO

Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization

69

NVlabs/edm2

EDM2 and Autoguidance -- Official PyTorch implementation

853

shinying/dmp

[CVPR 2024] Exploiting Diffusion Prior for Generalizable Dense Prediction

81

G-U-N/Diffusion-NPO

[ICLR 2025, AAAI 2026] official implementation of "Diffusion-NPO: Negative Preference Optimization…

39

InternScience/AdaptiveDiffusion

[NeurIPS'24] Training-Free Adaptive Diffusion with Bounded Difference Approximation Strategy

74

okotaku/diffengine

Diffusers training with mmengine

100

Languages

Python

97.4%

Shell

1.5%

Makefile

1.1%

HKUST-SAIL/NoiseAR

Official implementation of "NoiseAR: AutoRegressing Initial Noise Prior for Diffusion Models"

18

stars

14

commits

Python

primary language

Jun 3, 2025

updated

README

NoiseAR: AutoRegressing Initial Noise Prior for Diffusion Models

Zeming Li$^{*}$ Xiangyue Liu$^{{*}}$, Xiangyu Zhang, Ping Tan, Heung-Yeung Shum

Getting Started

1. Download Documents for Setting Up the Environment, Our Pretrained Models, and Eval Datasets (Compulsory)

bash download.sh

2. Set Up Your Development Environment

conda create --name noise_ar python=3.10.8 
conda activate noise_ar
make install

3. Run Inference

For example, to run inference using SDXL with our trained model:

python try.py

4. Run Evaluation Metrics

For example, run eval using SDXL with our trained model on DrawBench Evaluation Dataset:

torchrun eval.py exp.suffix=eval_sdxl_DrawBench exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/DrawBench
Click to expand for all of our evaluation commands 
# SDXL on DrawBench, GenEval, PickaPic
torchrun --master_port 13340 eval.py exp.suffix=eval_sdxl_DrawBench exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/DrawBench
torchrun --master_port 13341 eval.py exp.suffix=eval_sdxl_GenEval exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/GenEval
torchrun --master_port 13342 eval.py exp.suffix=eval_sdxl_PickaPic exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/PickaPic

# DreamShaper on DrawBench, GenEval, PickaPic
torchrun --master_port 13343 eval.py exp.suffix=eval_dreamshaper_DrawBench exp.pipeline=DreamShaper exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/DrawBench exp.cfg=3.5
torchrun --master_port 13344 eval.py exp.suffix=eval_dreamshaper_GenEval exp.pipeline=DreamShaper  exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/GenEval exp.cfg=5.0
torchrun --master_port 13345 eval.py exp.suffix=eval_dreamshaper_PickaPic exp.pipeline=DreamShaper  exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper/model.pth exp.val_data_path=data/PickaPic exp.cfg=5.0

# DiT on DrawBench, GenEval, PickaPic
torchrun --master_port 13346 eval.py exp.suffix=eval_dit_DrawBench exp.pipeline=DiT exp.pretrained_path=pretrained_models/dit/model.pth exp.val_data_path=data/DrawBench 
torchrun --master_port 13347 eval.py exp.suffix=eval_dit_GenEval exp.pipeline=DiT  exp.pretrained_path=pretrained_models/dit/model.pth exp.val_data_path=data/GenEval
torchrun --master_port 13348 eval.py exp.suffix=eval_dit_PickaPic exp.pipeline=DiT  exp.pretrained_path=pretrained_models/dit/model.pth exp.val_data_path=data/PickaPic 

# DPO of SDXL, DreamShaper, DiT on DrawBench
torchrun --master_port 13349 eval.py exp.suffix=eval_sdxl_DrawBench_DPO exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper_dpo/model.pth exp.val_data_path=data/DrawBench
torchrun --master_port 13350 eval.py exp.suffix=eval_dreamshaper_DrawBench_DPO exp.pipeline=DreamShaper exp.pretrained_path=pretrained_models/sdxl_and_dreamshaper_dpo/model.pth exp.val_data_path=data/DrawBench
torchrun --master_port 13351 eval.py exp.suffix=eval_dit_DrawBench_DPO exp.pipeline=DiT exp.pretrained_path=pretrained_models/dit_dpo/model.pth exp.val_data_path=data/DrawBench exp.cfg=5.0

🚧 Todo

  • Release the training code & data.
  • Release the training code & data for DPO.

📍 Citation

If you find this project useful for your research, please cite:

@misc{li2025noisearautoregressinginitialnoise,
      title={NoiseAR: AutoRegressing Initial Noise Prior for Diffusion Models}, 
      author={Zeming Li and Xiangyue Liu and Xiangyu Zhang and Ping Tan and Heung-Yeung Shum},
      year={2025},
      eprint={2506.01337},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2506.01337}, 
}

Similar repos

xie-lab-ml/Golden-Noise-for-Diffusion-Models

[ICCV2025] The code of our work "Golden Noise for Diffusion Models: A Learning Framework".

197

Klayand/Golden-Noise-for-Diffusion-Models

Inference code of "Golden Noise for Diffusion Models: A Learning Framework".

4

Kwai-Kolors/LPO

Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization

69

NVlabs/edm2

EDM2 and Autoguidance -- Official PyTorch implementation

853

shinying/dmp

[CVPR 2024] Exploiting Diffusion Prior for Generalizable Dense Prediction

81

G-U-N/Diffusion-NPO

[ICLR 2025, AAAI 2026] official implementation of "Diffusion-NPO: Negative Preference Optimization…

39

InternScience/AdaptiveDiffusion

[NeurIPS'24] Training-Free Adaptive Diffusion with Bounded Difference Approximation Strategy

74

okotaku/diffengine

Diffusers training with mmengine

100

Contributors

Lxiangyue

14 commits

Languages

Python

97.4%

Shell

1.5%

Makefile

1.1%