DFloat11 Compressed Model: OmniGen2/OmniGen2 Transformer
This is a DFloat11 losslessly compressed version of the original OmniGen2/OmniGen2 model. It reduces model size by 32% compared to the original BFloat16 model, while maintaining bit-identical outputs and supporting efficient GPU inference.
📊 Performance Comparison
| Metric | OmniGen2 (BFloat16) | OmniGen2 (DFloat11) |
|---|
| Model Size | 16.23 GB | 11.11 GB |
Peak GPU Memory
(1024×1024 image generation) | 18.41 GB | 14.36 GB |
Generation Time
(A100 GPU) | 25 seconds | 27 seconds |
🔧 How to Use
A complete usage guide is available in our GitHub repository (forked from the official OmniGen2 repository).
🔍 How It Works
We apply Huffman coding to losslessly compress the exponent bits of BFloat16 model weights, which are highly compressible (their 8 bits carry only ~2.6 bits of actual information). To enable fast inference, we implement a highly efficient CUDA kernel that performs on-the-fly weight decompression directly on the GPU.
The result is a model that is ~32% smaller, delivers bit-identical outputs, and achieves performance comparable to the original BFloat16 model.
Learn more in our
research paper.
📄 Learn More
