dingshizhe/vgtae-cc12m-latents

CC12M latents encoded with VGT-AE (448px) 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. Contents 1097 WebDataset… See the full description on the dataset page: https://huggingface.co/datasets/dingshizhe/vgtae-cc12m-latents.

<iframe src="https://silytics.si/embed/item/74bdfc4f-8796-4cd9-b044-a4b36aab7348" width="420" height="190" style="border:0" loading="lazy" title="Silytics entry"></iframe>

Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 18 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.