Arno Solin

While this is a very interesting paper, I was disappointed to see that the authors missed our extensive prior work on this topic. We have been thinking about this problem since 2023: - Ambient Diffusion: Learning Clean Distributions from Corrupted Data (NeurIPS 2023) - Consistent Diffusion Meets Tweedie: Training Exact Ambient Diffusion Models with Noisy Data (ICML 2024) - Does Generation Require Memorization? Creative Diffusion Models using Ambient Diffusion (ICML 2025) - Ambient Diffusion Posterior Sampling: Solving Inverse Problems with Diffusion Models Trained on Corrupted Data (ICLR 2025) - Ambient Diffusion Omni: Training Good Models with Bad Data (NeurIPS spotlight 2025) ... and a lot more works from our group and others, e.g. - Measurement Score-Based Diffusion Model (ICLR 2026). While I fully understand that it is hard to keep up with the literature these days, it still feels very sad not to receive proper credit for our contributions to this problem.

Training a diffusion model has always been synonymous with one idea: add some noise to an image, then learn to remove it. Since we know where we started and where we ended up, the natural thing to do is to ask the model to recover the signal everywhere. But is it really needed?