Two recent arXiv preprints draw attention to different weak points in the lifecycle of diffusion models. The first, 'The critical slowing down in training diffusion models,' focuses on the training phase. Its abstract places the work in the broader history of computational sampling and notes that machine-learning approaches have enabled major advances even though their behavior remains poorly understood. The title identifies 'critical slowing down' as a specific phenomenon worth investigating, though the abstract excerpt does not explain the underlying mechanism.
The second paper, 'Quantization-Aware Kalman Estimation for Diffusion Sampling,' is concerned with deployment. It describes quantization as a practical way to reduce memory and computation, but cautions that aggressive compression can cause quantized outputs to deviate substantially from their full-precision counterparts. The title indicates a Kalman-estimation approach that accounts for quantization.
The two papers are largely complementary. One examines what happens while a diffusion model is being trained; the other examines what happens when a trained model is compressed for real-world use. They do not directly overlap in method or subject matter.
What they share is a common underlying concern: diffusion models are powerful but not yet fully understood. Whether the issue is a training dynamic like critical slowing down or a compression artifact like quantization drift, both papers point to the need for more robust ways to control these models.