DiffBMP: Differentiable Rendering with Bitmap Primitives

Accepted to CVPR 2026

Authors

Seongmin Hong^1,*, Junghun James Kim^2,*, Daehyeop Kim³, Insoo Chung³, Se Young Chun^1,2,3,†

¹ INMC, ² IPAI, ³ Dept. of ECE, Seoul National University, Republic of Korea

^* Co-first authors    ^† Corresponding author

Overview

DiffBMP is a fast differentiable rendering framework for creating vector art with arbitrary bitmap primitives. Unlike traditional vector graphics that rely on geometric primitives, DiffBMP enables the use of any image as a primitive, opening new possibilities for artistic expression and image synthesis.

🚀 Installation

You have two options for installation:

Option 1: PyPI Installation (Linux/WSL Only - Simple & Fast)

Best for users who want to quickly use DiffBMP without building from source.

# Install PyTorch first (if not already installed)
# Note: You can use any CUDA version compatible with your system (cu118, cu121, cu124, etc.)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

# Install pydiffbmp
pip install pydiffbmp

# Install system dependencies
sudo apt-get install poppler-utils

# Install additional Python dependencies (if needed)
pip install tqdm nvidia-ml-py fonttools svgwrite svgpathtools scour cairosvg psd-tools piq moviepy pyvista pandas scikit-image scipy

System Requirements

• Python: 3.8, 3.9, 3.10, or 3.11
• Platform: Linux x86_64 (Windows/macOS not supported for PyPI installation)
• GPU: NVIDIA GPU with CUDA capability 8.6+ (RTX 3090, A100, RTX 4090, L40S, etc.)
• CUDA Driver: 11.8+ (CUDA toolkit NOT required for PyPI installation)
• PyTorch: 1.13.0+

⚠️ Note: Pre-built binaries are currently available for Linux only. Windows users should use WSL2 or build from source.

Option 2: Build from Source (All Platforms - Flexible)

Best for developers, customization, or if you need to build on Windows/macOS.

# Clone the repository
git clone https://github.com/smhongok/diffbmp.git
cd diffbmp

# Install dependencies
pip install -r requirements.txt

# Build CUDA extensions
cd cuda_tile_rasterizer
python setup.py build_ext --inplace
python setup_fp16.py build_ext --inplace
python setup_psd_export.py build_ext --inplace
cd ..

# Install in editable mode
pip install -e .

Build Requirements:

• CUDA Toolkit 12.3+ (recommended)
• C++ compiler (GCC 9.4.0+ recommended)
• PyTorch 1.13.0+

For detailed build instructions and troubleshooting, see DEVELOPMENT.md.

Running the Original Research Code

The original research implementation is available in the GitHub repository:

# Clone the repository
git clone https://github.com/smhongok/diffbmp.git
cd diffbmp

# Run with configuration
python main.py --config configs/default.json

For more details on research code, evaluation scripts, and advanced features, see DEVELOPMENT.md.

📚 Citation

If you use DiffBMP in your research, please cite our paper:

@misc{hong2026diffbmp,
    author    = {Hong, Seongmin and Kim, Junghun James and Kim, Daehyeop and Chung, Insoo and Chun, Se Young},
    title     = {DiffBMP: Differentiable Rendering with Bitmap Primitives},
    arxiv     = {2602.22625},
    year      = {2026}
}

🙏 Acknowledgments

This work was inspired by and builds upon several excellent open-source projects:

• DiffVG
• gsplat

We thank the authors of these projects for their contributions to the community.

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.

For questions or issues, please open an issue on GitHub.