NeurIPS 2025
Install other pip packages via pip install -r requirements.txt.
Download the pre-trained model and put them under data/pretrain.
We use Matterport3D in our experiments and follow PanFusion for data preparation. The data should be organized as follows.
data
├── Matterport3D
│ ├── mp3d_skybox
│ │ ├── 1LXtFkjw3qL
│ │ ├── matterport_skybox_images
│ │ └── matterport_stitched_images
│ │ ├── train.npy
│ │ └── test.npy
│ └── caption
│ ├── 1LXtFkjw3qL_0b22fa63d0f54a529c525afbf2e8bb25.txt
└── pretrain
Download the checkpoint and put it under data/pretrain/ckpt/par_w1024.pt.
accelerate launch scripts/demo.py --ptr data/pretrain/nova-d48w1024-sdxl1024 --ckpt data/pretrain/ckpt/par_w1024.pt --prompt=[Prompt]
The codes in scripts is launched by accelerate. The images and captions are specified by path and textpath, respectively.
accelerate launch scripts/train.py configs/cfg.yaml ptr=[ptr] path=[path] textpath=[textpath] env.o=[project path]
accelerate launch scripts/train.py configs/cfg.yaml ptr=[ptr] path=[path] textpath=[textpath] env.o=[project path] eval_only=true
If you find our work interesting, please kindly consider citing our paper:
@article{wang2025conditional,
title={Conditional Panoramic Image Generation via Masked Autoregressive Modeling},
author={Wang, Chaoyang and Li, Xiangtai and Qi, Lu and Lin, Xiaofan and Bai, Jinbin and Zhou, Qianyu and Tong, Yunhai},
journal={arXiv preprint arXiv:2505.16862},
year={2025}
}Apache License 2.0