SphereDiff: Tuning-free 360° Static and Dynamic Panorama Generation via Spherical Latent Representation

Minho Park*, Taewoong Kang*, Jooyeol Yun, Sungwon Hwang and Jaegul Choo
Korea Advanced Institute of Science and Technology (KAIST)
AAAI 2026 (Oral). (* indicate equal contribution)

🌐 Overview

SphereDiff enables tuning-free generation of 360° panoramic images and videos using pretrained diffusion models.
Unlike ERP-based methods, SphereDiff defines a spherical latent representation to maintain consistent quality across all viewing directions.

Key features:

• 🌍 Spherical latent representation for distortion-free 360° generation
• 🌀 Supports both static and dynamic panoramas
• ⚙️ Plug-and-play with existing pretrained diffusion models
• 💡 No additional fine-tuning required

🎥 Demo

Generated 360° Static and Dynamic Panoramas with Diverse Diffusion Backbones.

🚀 Example Usage

📦 Installation

conda create -n spherediff python=3.10

# install pytorch according to your cuda version
pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu128

# then, install requirements
pip install -r requirements.txt

Run SphereDiff without any additional training!

task_name="StaticWallpapers"
pipeline_name="SphericalFluxPipeline"
default_config="
pipeline_cls=${pipeline_name}
pretrained_model_name_or_path=black-forest-labs/FLUX.1-dev
variant=None
mixed_precision=bf16
enable_model_cpu_offload=False
call_kwargs.n_spherical_points=26500
"
subdir="my_subdir"
txt_name="ruins"
prompt_txt_path="data/prompts/${txt_name}.txt"
save_path="./outputs/${task_name}/${pipeline_name}/${subdir}/${txt_name}"

python generate_static_wallpaper.py --config_add ${default_config} call_kwargs.prompt_txt_path=${prompt_txt_path} save_path=${save_path} ;

See ./scripts/run_spherediff.md for full options.

🚧 Planned Updates

• Update the project page and arXiv link
• Release the base code for static & live wallpaper generation
• Release the code for foreground–background generation

📚 Citation

@article{park2025spherediff,
  title={SphereDiff: Tuning-free Omnidirectional Panoramic Image and Video Generation via Spherical Latent Representation},
  author={Park, Minho and Kang, Taewoong and Yun, Jooyeol and Hwang, Sungwon and Choo, Jaegul},
  journal={arXiv preprint arXiv:2504.14396},
  year={2025}
}