Pre-computed posterior means (no variance/std) from the Stable Diffusion VAE (stabilityai/sd-vae-ft-mse) for the full ImageNet-1K training set at 256×256 resolution, stored as Parquet shards. Each example includes latents for both the original and horizontally flipped image, enabling flip augmentation without re-encoding at training time.
