Answers are grounded in this paper and cite their sources.
Investigating likelihood reliability in generative models using feature representations, suggesting avenues for improvement.
Normalising flows are a flexible class of generative models that provide exact likelihoods and are often trained through maximum likelihood estimation. Recent work suggests that discrete‐step flow models trained in this way can assign undesirably high likelihood to out‐of‐distribution image data, bringing their reliability for applications where likelihoods are important (e.g., outlier detection) into question. Continuous‐time normalising flows trained with the conditional flow matching objective (CFM models) also provide unreliable likelihoods, and we investigate whether training them on various feature representations can lead to more reliable likelihoods. We consider features from a pretrained classifier, features from a pretrained perceptual autoencoder and features from an autoencoder trained from scratch with a simple pixel‐based reconstruction loss, and compare their effects on CFM model likelihoods on various in‐ and out‐of‐distribution sets. Autoencoder‐based features are of particular interest as the presence of a decoder preserves the ability to generate images. We find that training CFM models on feature representations can lead to improvements in likelihood reliability, but only for certain datasets and certain parameterisations of the feature space, at a cost in sample quality. Further investigation suggests possible links between likelihood reliability and geometric characteristics of the data and the feature space, and opens avenues for future work.
No takes yet. Share an insight, caveat, or question.
Josias et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: