core

Transition Matching: Scalable and Flexible Generative Modeling

Neta Shaul, Uriel Singer, Itai Gat, Yaron Lipman · Meta AI (FAIR) · 2025-06 · arXiv:2506.23589

Why it matters

Recasts generation as learning discrete-time Markov transition kernels over a continuous state, which unifies flow matching and continuous-token autoregression in one formulation instead of treating them as rival camps. Its fully-causal variant is reported as the first such model to match or beat flow-based methods on text-to-image.

What this paper does

read: full text

Before this

Diffusion and flow matching had an extensively investigated design space that the authors argue offered limited further headroom. Continuous-token autoregressive generation was emerging as a way to unify text and media generation, but with less theoretical grounding and generally weaker or slower quality than flow-based methods, and the two paradigms were treated as separate camps.

The problem

Whether diffusion and flow matching on one side, and continuous autoregressive generation on the other, can be described as instances of a single more general framework, and whether the extra design freedom that framework allows can improve on both.

The idea

Recast generation as learning discrete-time Markov transition kernels between consecutive states over a small number of steps, rather than committing to either a continuous ODE or SDE or a strict token-by-token autoregression. Because the kernel at each step can be an expressive, non-deterministic distribution rather than a fixed deterministic velocity, the same formulation covers a flow-matching-like variant and two autoregressive variants that differ only in how much conditioning history each transition sees.

How it works

Generation happens in the latent space of an SDXL VAE, a 32 by 32 by 4 grid for 256-pixel images, not in pixels or a semantic representation space. Three variants share a common backbone, either a 1.7B-parameter DiT or an LLM, plus a small 40M-parameter MLP head, and differ in what the head's kernel predicts and conditions on. Difference Transition Matching, DTM, generalizes flow matching to discrete time by learning the posterior distribution of Y equals X_T minus X_0, a stochastic drift-like difference field between the noisy and clean states, and it generates all tokens of a step independently and non-causally. Autoregressive Transition Matching, ARTM, learns the same style of per-token transition kernel but conditions each token's prediction on previously generated tokens within the same state, making it partially causal. Full History Transition Matching, FHTM, additionally conditions on the entire trajectory of past states, giving a fully causal backbone compatible with teacher forcing and with integration into large language model architectures. All variants are trained with a flow-matching-style loss on their kernel's drift field, with text conditioning dropped at probability 0.15 for classifier-free guidance.

Evidence

Trained for 500K iterations at batch size 2048 on 350M licensed Shutterstock image-caption pairs, DTM reaches CLIPScore 26.84, PickScore 21.18 and ImageReward 0.532 on PartiPrompts using classifier-free guidance at scale 6.5 and only 16 backbone forward passes, against a flow-matching baseline's CLIPScore 25.97, PickScore 21.04 and ImageReward 0.233 using 128 forward passes, roughly a sevenfold reduction in sampling cost at matched or better quality. FHTM-3, the fully causal variant, reaches CLIPScore 27.00 and GenEval overall 0.52 against the flow-matching baseline's GenEval 0.47, which the paper presents as the first fully causal model to match or surpass flow-based methods on text-to-image in a continuous domain. ARTM and FHTM need substantially more backbone forwards than DTM, for example 3 times 256 against DTM's 32, so neither shares DTM's speed advantage.

Limitations

The authors concede that DTM's kernel generates every token within a step independently, which limits the kernel's expressive power. They also concede that the autoregressive kernels exploit the linear structure of a linear noising process and collapse to a degenerate function under it, which is why ARTM and FHTM instead use an independent-per-step noising process. They note limited exploration of time schedulers and distillation for the new kernels.

Why it matters

Putting diffusion, flow matching and continuous-token autoregression inside one Markov-transition formulation turns what looked like separate research programs into different settings of a single design space. The fully causal FHTM variant gives a concrete path to folding continuous image generation into the same causal architectures used for language modeling.

Abstract, in the authors' own words

Diffusion and flow matching models have significantly advanced media generation, yet their design space is well-explored, somewhat limiting further improvements. Concurrently, autoregressive (AR) models, particularly those generating continuous tokens, have emerged as a promising direction for unifying text and media generation. This paper introduces Transition Matching (TM), a novel discrete-time, continuous-state generative paradigm that unifies and advances both diffusion/flow models and continuous AR generation. TM decomposes complex generation tasks into simpler Markov transitions, allowing for expressive non-deterministic probability transition kernels and arbitrary non-continuous supervision processes, thereby unlocking new flexible design avenues. We explore these choices through three TM variants: (i) Difference Transition Matching (DTM), which generalizes flow matching to discrete-time by directly learning transition probabilities, yielding state-of-the-art image quality and text adherence as well as improved sampling efficiency. (ii) Autoregressive Transition Matching (ARTM) and (iii) Full History Transition Matching (FHTM) are partially and fully causal models, respectively, that generalize continuous AR methods. They achieve continuous causal AR generation quality comparable to non-causal approaches and potentially enable seamless integration with existing AR text generation techniques. Notably, FHTM is the first fully causal model to match or surpass the performance of flow-based methods on text-to-image task in continuous domains. We demonstrate these contributions through a rigorous large-scale comparison of TM variants and relevant baselines, maintaining a fixed architecture, training data, and hyperparameters.

Research line

Transition matchingemerging

Design-axis choices

Objective

Transition matching

Method note — the shared flow-matching interpolation

Every flow-matching / rectified-flow paper in this atlas trains toward a straight-line path between a noise sample x₀ and a data sample x₁:

x_t = (1-t)\,x_0 + t\,x_1, \qquad v_\theta(x_t, t) \approx x_1 - x_0

Builds on

Nothing recorded yet.

Built on by

Nothing recorded yet.