v0.1.0 — Initial release

Initial release of diffusers-dmse.

This package provides DMSEScheduler, a lightweight extension of Hugging Face diffusers that implements a deterministic reverse diffusion process. By removing stochastic resampling, the scheduler follows the posterior mean trajectory and converges to the conditional mean estimator (CME), which is optimal in terms of mean squared error.

Highlights:

• Drop-in replacement for DDPMScheduler
• Deterministic reverse process (no sampling noise)
• SNR-based initialization via init_step()
• Compatible with pretrained diffusers pipelines

Use cases:

• Denoising from noisy observations with known SNR
• Deterministic diffusion trajectories
• Reproducible evaluation and research experiments

Related work:

• "On the Asymptotic Mean Square Error Optimality of Diffusion Models" (AISTATS 2025)
  https://arxiv.org/abs/2403.02957

Installation:
pip install diffusers-dmse

Repository:
https://github.com/benediktfesl/diffusers-MSEopt