CDM pushes diffusion distillation into continuous time

Hot take: this is a careful incremental advance on diffusion distillation rather than a brand-new paradigm, but the continuous-time framing looks genuinely useful if the reported quality holds up outside the paper’s evaluation set.

• –The core idea is to remove the brittleness of sparse timestep anchoring and supervise along a continuous schedule instead.
• –The method also adds off-trajectory matching via the student’s velocity field, which is the more interesting part technically.
• –The reported results are strong for a distillation paper: better fidelity metrics at 4 NFE than DMD2/D-DMD on the showcased setups.
• –The lack of GANs or reward models is practical, since those extras often make diffusion acceleration pipelines harder to train and maintain.
• –Main caveat: this is still a paper-level result on specific architectures and benchmarks, so real-world generalization is the thing to watch.