Koopman Observers for Diffusion Acceleration: Correcting Feature Forecasts with Shallow Measurements

Key points

• Feature caching accelerates diffusion sampling by replacing expensive network evaluations with predictions from previously computed activations.1
• Using calibration trajectories, we identify finite-dimensional, time-dependent Koopman approximations that jointly describe the increments of shallow and deep network features.1
• This formulation enables controlled comparisons of temporal prediction and observation correction.1
• The observer achieves $1.89\times$ and $1.85\times$ measured speedups over DDIM-50, supporting improved reference-sampler fidelity without retraining the denoiser.1

Sources (1)

• [1]Koopman Observers for Diffusion Acceleration: Correcting Feature Forecasts with Shallow Measurements
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 7, 04:35 PM

  We introduce an observation-corrected Koopman framework for accelerating frozen diffusion models.

  Feature caching accelerates diffusion sampling by replacing expensive network evaluations with predictions from previously computed activations.

Extractive summary: sentences quoted from the sources.