Flow Matching

Flow matching models generate images by learning a continuous probability flow (ODE trajectory) from the noise distribution to the data distribution, and are generally more stable to train and require fewer inference steps.

The idea of flow matching is to learn a time-dependent vector field v_t(x) such that the ordinary differential equation driven by this vector field can transport a simple reference distribution to the target data distribution.

Acceleration solutions such as consistency models and flow matching are narrowing the computational overhead gap of multi-step inference in diffusion models.

In the training stage, the method optimizes a signal-specific 1D Transformer denoising network via flow matching so that it directly predicts clean signal samples.

Flow-matching action experts (such as π0.5) model action generation as a continuous transformation process from noise to actions, producing smooth action sequences through iterative denoising and suiting high-dimensional continuous control; autoregressive methods (such as π0-FAST) discretize actions into tokens and predict them token by token.

In recent years, generation methods based on neural codecs and diffusion/flow matching have brought synthetic speech much closer to real people in prosody, breathiness, and timbre detail.

After removing the acoustic tokenizer, the model must model directly in a continuous acoustic space, typically relying on continuous generation paradigms such as diffusion models or flow matching, which shortens the inference pipeline.

The Velocity Ambiguity Profile (VAP) is a time-resolved Bayesian risk diagnostic tool for revealing reducible and irreducible error in flow matching training.

VAP turns the averaged flow matching objective into a time-resolved diagnostic tool, helping locate ambiguity-limited and data-limited regions.

Flow matching is a technique for training continuous-time normalizing flows that learns a transformation by matching the vector field along a probability path, avoiding the computational cost of simulating ordinary differential equations during training.

The 'noise stage' or 'early flow steps' of flow matching refers to the first few steps of integration where t is close to 0 and the sample is still close to pure noise; at this stage the image structure information has not yet taken shape, and this is exactly the time window in which CFG behavior becomes anomalous.

Classifier-Free Guidance (CFG) is one of the most mainstream guidance techniques in diffusion models and flow matching models, discarding conditions with a certain probability during training and, during sampling, subtracting the unconditional prediction from the conditional one via a formula and amplifying the guidance strength.

In stage three, SUPRA-TEX adopts an interleaved flow transformer, using the observation latents as fixed read-only anchors and alternating between updating unobserved regions and updating the entire surface, ultimately generating complete material latents.

PathFM uses flow matching to generate complete future paths.

Fixed-point score matching is a unified framework proposed by Dennis Blessing that unifies diffusion sampling algorithms such as marginal shooting, Schrödinger bridge, flow matching, and bridge matching sampling into the same underlying principle.

A study published on arXiv (arXiv:2610.10823v1) offers a different answer from the mainstream explanation regarding why velocity scaling in flow matching is effective.

Compared to diffusion models, flow matching has a more direct training objective and straighter trajectories, theoretically enabling high-quality results with fewer sampling steps (NFE).

The framework uses a flow matching model during training to learn the joint distribution of adjacent axial slice triplets.

Flow Matching can be viewed as a special case of the deterministic version of stochastic interpolators.

The arXiv paper (arXiv:2610.02663) systematically studies the generalization properties of score-matching-based diffusion models/flow matching models on intrinsically low-dimensional data.