Making normalizing flows competitive for image generation

Normalizing flows can compute exact likelihoods and generate images deterministically, but struggle with high-dimensional data compared to diffusion models. The problem: flows must learn a single invertible path through the entire feature space, which wastes capacity on noisy, overcomplete representations. SRC-Flow inserts a compression step that squeezes redundant features into a compact semantic space before flow modeling, then reconstructs images through a frozen decoder. On ImageNet at 256×256 and 512×512, this achieves state-of-the-art flow-based generation (gFID 1.65 and 2.07) while preserving exact likelihood and deterministic sampling—useful for applications requiring provable density estimates or reproducible generation.