Latent Diffusion for Missing Data

A new study proposes shifting diffusion-based imputation from pixel space to a learned latent space, reporting that the approach remains stable even when half the training data is missing [1]. The work, posted to arXiv on 27 May 2026 by Alberte Heering Estad, introduces a two-stage framework designed to handle missing-completely-at-random corruption [1]. A robust VAE-based imputer first learns compact semantic features from incomplete observations, and a diffusion model is then trained in the resulting latent space [2]. The authors performed a controlled comparison against pixel-space diffusion models under identical incomplete-data conditions [2]. They found that the latent diffusion model maintains high sample quality and remains stable up to 50% missingness, while pixel-space diffusion degrades progressively as missingness increases [2]. For downstream imputation tasks, latent diffusion also achieved consistently better performance than its pixel-space counterpart [2]. The paper argues that latent-space modeling mitigates artifact amplification from zero-imputed inputs and provides a more robust generative prior for incomplete-data learning [2]. Diffusion models belong to a broader class of generative AI systems that learn underlying patterns in training data and generate new samples in response to input [3]. The prevalence of such tools has grown sharply since the AI boom of the 2020s, driven by advances in deep neural networks and large language models [3]. Within this landscape, latent-space representations are already used in other domains: vision-language-action models, for instance, translate image observations and natural language descriptions into a distribution within a latent space before decoding actions [4]. The concept of latent variables has a longer statistical lineage, including the expectation–maximization algorithm, which iteratively estimates parameters in models with unobserved latent variables [5]. The submission file size is 698 KB [1]. The findings position latent diffusion as a practical alternative for missing-data problems, particularly in settings where training sets are heavily incomplete [2].