STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis

TL;DR

STARFlow is a Transformer-parameterized autoregressive normalizing flow that operates on the latent space of a pretrained VAE, scaled to text-conditional high-resolution image synthesis. Its central architectural move is a deep–shallow decomposition: a small stack of shallow alternating-direction AF blocks reshapes the latent locally, then a single deep causal-Transformer flow block carries the bulk of capacity for global semantic modeling. The work also revisits guidance, sampling, and decoder fine-tuning for NFs and shows that normalizing flows — long considered uncompetitive with diffusion — can match strong T2I diffusion baselines on standard benchmarks, providing the building block that STARFlow-V then ports to video STARFlow-V: End-to-End Video Generative Modeling with Normalizing Flows.

Key claims

Stacking AF blocks with alternating left-to-right / right-to-left causal masks is sufficient for universal density modeling — the affine form of each block plus mask alternation captures bidirectional dependencies without breaking invertibility [§2.2, Eq. 2.2].
A deep–shallow split — many shallow alternating-direction blocks doing local reshaping plus one deep causal-Transformer block carrying most parameters — is the configuration that makes Transformer-parameterized AF scale; allocating capacity to the deep block alone for semantic modeling beats balanced splits [§3].
Training in the latent space of a pretrained VAE (SD-VAE) at 256² rather than in pixel space is what makes high-resolution NF tractable; the flow models a continuous latent rather than discrete pixels [§3, Architecture].
Small-noise data augmentation during training (σ-smoothed density) is necessary for NF training stability, but its side-effect — slightly noisy samples — must be removed post-hoc; STARFlow fine-tunes the VAE decoder with a GAN objective to denoise noisy latents at decode time [§3.2].
Block-wise Jacobi iteration converts the AF inverse from strictly sequential token-by-token sampling into a parallelizable fixed-point solve, with a scale-normalized residual stopping criterion (default 1e-3) [§3.3].
The model is classifier-free-guidance-compatible: guided affine parameters (μ̂, σ̂) can be computed from conditional and unconditional forward passes and substituted into the inverse update [§3.3].

Method

STARFlow takes a 256×256 image, encodes it via SD-VAE, and models the latent as an autoregressive normalizing flow. Each AF block applies an affine transform y = s ⊙ x + t where (s, t) are produced by a causal Transformer over the masked input; blocks alternate left-to-right and right-to-left to give effectively bidirectional context while remaining invertible. The architecture stacks several shallow such blocks followed by one large deep block (the model is 3B parameters total, with most of the capacity in the deep block). Text conditioning is via T5-XL, and RoPE is used for positional encoding. Training maximizes exact log-likelihood under the change-of-variables formula with σ-noised inputs; the VAE decoder is subsequently fine-tuned with a GAN loss so that decoding remains clean when fed slightly noisy latent samples. Sampling inverts the flow block-wise via Jacobi iteration with KV caching across completed blocks.

Results

The paper is a NeurIPS 2025 Spotlight and reports text-to-image generation quality competitive with SDXL-class diffusion baselines on standard FID/CLIP benchmarks. Headline numbers are not extractable from the body retrieved here, but the central empirical claim — that a Transformer-parameterized AF in latent space matches strong T2I diffusion at 256² — is what motivates the follow-up video work; STARFlow-V cites STARFlow as the foundation on which the global–local video architecture is built STARFlow-V: End-to-End Video Generative Modeling with Normalizing Flows §1.

Why it’s interesting

STARFlow is the architectural foundation that made normalizing flows’ 2025 comeback possible at the high-resolution image domain: TARFlow, STARFlow, Farmer, and now STARFlow-V are all latent-space Transformer-parameterized AF variants, and STARFlow is the one that first hit T2I quality at scale. From a Luma research perspective the interesting bits are the deep–shallow capacity allocation and the small-noise + decoder-fine-tune training recipe — both reusable in any latent-space generator. The model is also a clean counterpoint to the Diffusion Distillation lineage: rather than training a diffusion model and then compressing it down to few steps, STARFlow uses an end-to-end likelihood objective from the start, and reaches comparable image quality without a distillation stage. Pairs naturally with Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (diffusion-as-pretraining-for-one-step generators), which approaches the same “no iterative sampling at inference” goal from the diffusion side.