Hi-MAR

🖥️ GitHub    ｜    🌐 Project Page    |   🤗 Hugging Face   |    📑 Paper    |    📖 PDF   

Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025)

This is the official repository for the Paper "Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots"

Overview

We present a Hierarchical Masked Autoregressive models (Hi-MAR) that pivot on low-resolution image tokens to trigger hierarchical autoregressive modeling in a multi-phase manner.

🔍 What We're Working to Solve?

• Incapable of utilizing global context in early-stage predictions of the next-token paradigm
• Training-inference discrepancy across multi-scale predictions
• Suboptimal multi-scale probability distribution modeling
• Lack of global information in the denoising process of the MLP-based Diffusion head

🔥 Updates

• [2025.05.22] Upload inference code and pretrained class-conditional Hi-MAR models trained on ImageNet 256x256.

🏃🏼 Inference

Environment Requirement

Clone the repo:

git clone https://github.com/HiDream-ai/himar.git
cd himar

Install dependencies:

conda env create -f environment.yaml

conda activate himar

Model Download

Download VAE from the link in the MAR Github.

You can download our pre-trained Hi-MAR models directly from the links provided here.

Models	FID-50K	Inception Score	#params
Hi-MAR-B	1.93	293.0	244M
Hi-MAR-L	1.66	322.3	529M
Hi-MAR-H	1.52	322.78	1090M

Evaluation

Evaluate Hi-MAR-B on ImageNet256x256:

torchrun --nproc_per_node=8 --nnodes=1 main_himar.py --img_size 256 --vae_path /path/to/vae --vae_embed_dim 16 --vae_stride 16 --patch_size 1 --model himar_base --diffloss_d 6 --diffloss_w 1024 --output_dir ./himar_base_test --resume /path/to/Hi-MAR-B  --num_images 50000 --num_iter 4 --cfg 2.5 --re_cfg 2.7 --cfg_schedule linear --cond_scale 8 --cond_dim 16 --two_diffloss --global_dm --gdm_d 6 --gdm_w 512 --eval_bsz 256 --load_epoch -1 --head 8 --ratio 4 --cos --evaluate

Evaluate Hi-MAR-L on ImageNet256x256:

torchrun --nproc_per_node=8 --nnodes=1 main_himar.py --img_size 256 --vae_path /path/to/vae --vae_embed_dim 16 --vae_stride 16 --patch_size 1 --model himar_large --diffloss_d 8 --diffloss_w 1280 --output_dir ./himar_large_test --resume /path/to/Hi-MAR-L  --num_images 50000 --num_iter 4 --cfg 3.5 --re_cfg 3.5 --cfg_schedule linear --cond_scale 8 --cond_dim 16 --two_diffloss --global_dm --gdm_d 8 --gdm_w 512 --eval_bsz 256 --load_epoch -1 --head 8 --ratio 4 --cos --evaluate

Evaluate Hi-MAR-H on ImageNet256x256:

torchrun --nproc_per_node=8 --nnodes=1 main_himar.py --img_size 256 --vae_path /path/to/vae --vae_embed_dim 16 --vae_stride 16 --patch_size 1 --model himar_huge --diffloss_d 12 --diffloss_w 1536 --output_dir ./himar_huge_test --resume /path/to/Hi-MAR-H  --num_images 50000 --num_iter 12 --cfg 3.2 --re_cfg 5.5 --cfg_schedule linear --cond_scale 8 --cond_dim 16 --two_diffloss --global_dm --gdm_d 12 --gdm_w 768 --eval_bsz 256 --load_epoch -1 --head 12 --ratio 4 --cos --evaluate

🌟 Star and Citation

If you find our work helpful for your research, please consider giving a star⭐ on this repository and citing our work.

💖 Acknowledgement

Thanks to the contribution of MAR