Abstract
Directly modeling the explicit likelihood of the raw data distribution is keytopic in the machine learning area, which achieves the scaling successes inLarge Language Models by autoregressive modeling. However, continuous ARmodeling over visual pixel data suffer from extremely long sequences andhigh-dimensional spaces. In this paper, we present FARMER, a novel end-to-endgenerative framework that unifies Normalizing Flows (NF) and Autoregressive(AR) models for tractable likelihood estimation and high-quality imagesynthesis directly from raw pixels. FARMER employs an invertible autoregressiveflow to transform images into latent sequences, whose distribution is modeledimplicitly by an autoregressive model. To address the redundancy and complexityin pixel-level modeling, we propose a self-supervised dimension reductionscheme that partitions NF latent channels into informative and redundantgroups, enabling more effective and efficient AR modeling. Furthermore, wedesign a one-step distillation scheme to significantly accelerate inferencespeed and introduce a resampling-based classifier-free guidance algorithm toboost image generation quality. Extensive experiments demonstrate that FARMERachieves competitive performance compared to existing pixel-based generativemodels while providing exact likelihoods and scalable training.