Anh Nguyen, Ngan Nguyen, Duc Vu, Trung Dao, Viet Nguyen, Quan Dao, Kien Nguyen, Chi Tran et al.
This paper introduces 'Cross-Space Distillation' and a lightweight bridge interface to enable knowledge transfer between teacher and student diffusion models with different latent spaces.
Distilling high-capacity modern teachers (e.g., SD 3.5, Flux) into compact students (e.g., SD 1.5) is hindered by the critical assumption that both models must share the same latent space, which differs in resolution and VAE parameterization.
The authors propose a lightweight 'Bridge' interface that maps student latents into the teacher's space without modifying the student backbone. It combines a frozen student VAE decoder with a compact learnable projector, trained for latent reconstruction and attention fidelity.
The Bridge significantly improves the performance of a compact SD 1.5 student (e.g., HPSv3 from 5.4 to 9.4) while preserving one-step inference and ecosystem compatibility. This demonstrates that heterogeneous large teachers can be distilled into efficient, deployable backbones via a lightweight latent-space interface.