ImageNet.3GB: Compressed ImageNet using Google SANA’s VAE
This dataset provides a compressed version of ImageNet using Google SANA’s VAE
(mit-han-lab/dc-ae-f32c32-sana-1.1-diffusers)
and 8-bit quantization, reducing the size to just 3GB while preserving high-quality reconstructions.
- Original ImageNet (256x256 JPEG): 16GB
- Float32 latent representation: 12GB
- This dataset (8-bit latents): 3GB
Usage
For encoding, decoding, and more details, visit the GitHub repository:
rotem154154/ImageNet.3GB