Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models

Specifically, we establish an information bottleneck by regularizing the language representation space, achieved through freezing the pretrained T2I diffusion decoder.