FARMER combines an invertible autoregressive flow, which can map images to and from latent space without information loss, with an autoregressive Transformer that models each latent token’s distribution using Gaussian mixtures, providing exact likelihoods in a single-stage, pixel-level generator. It introduces three key innovations: - Self-supervised channel split separates latent features into informative (Zᴵ) and redundant (Zᴿ) groups, efficiently modeling structure and detail. - Resampling-based classifier-free guidance (CFG) improves image quality and enables controllable sampling. - One-step AF distillation accelerates generation by replacing slow sequential reversal with a single fast reverse step. On ImageNet-256 (class-conditional, 50k samples), the 1.9B patch-8 model achieves FID 3.60 / IS 269.21 / Prec 0.81 / Rec 0.51. After +60 epochs, inference becomes 22× faster for AF reverse (0.1689s → 0.0076s per image) and about 4× faster overall (0.2189s → 0.0567s per image). It outperforms JetFormer 2.8B (FID 6.64) and TARFlow p8 (4.69), and is competitive with STARFlow p8. STARFlow’s decoder-finetuned variant (FID 2.40) remains stronger but uses a multi-stage setup.