ProtoFlow combines a vector‑quantized autoencoder with a prototype‑prior flow matching scheme to achieve efficient multivariate time‑series forecasting. First, a VQ‑AE encodes the raw series into a discrete latent space, yielding a set of codebook entries. Instead of initializing the flow from a generic Gaussian, the learned codebook is used to construct a structured prototype prior. Conditioned on historical observations, a DiT‑based rectified flow transports samples from this prototype distribution to future latent representations. By avoiding step‑by‑step autoregressive token generation, ProtoFlow eliminates exposure bias and the training‑inference mismatch, while accelerating convergence.
Across several public benchmark datasets, ProtoFlow consistently outperforms traditional diffusion models and VQ‑based autoregressive baselines in forecasting accuracy, yet retains the efficiency of a single‑step inference. Experiments demonstrate that the prototype prior leads to faster convergence and robust performance in high‑dimensional settings.
In summary, ProtoFlow leverages the learned discrete prototypes as an information‑rich prior, providing a concise and effective pathway for non‑autoregressive latent forecasting.
Review: The paper ingeniously repurposes the VQ codebook as a flow‑matching prior, addressing autoregressive exposure bias while preserving the flexibility of generative models. This approach merits further exploration in broader sequential generation tasks.