Pre-encoded CC12M
images in the VGT-AE latent space, so text-to-image training can skip the
encoder entirely.
These are not DC-AE latents. VGT-AE is a hybrid codec — a fine-tuned
Qwen2.5-VL ViT as the encoder, a DC-AE decoder at sampling time. The tensor shape
happens to match DC-AE f32c32 (32, 16, 16), but the space is completely
different; mixing the two silently trains a model against noise.
1097 WebDataset… See the full description on the dataset page:
https://huggingface.co/datasets/dingshizhe/vgtae-cc12m-latents.
