Dataset Condensation Atlas

P4 · Generative priors · since 2026 · ascendant

Autoregressive and flow-based synthesis

The generative-prior idea carried to visual autoregressive (next-scale) models and flow-matching generators.

The bet every member shares

The structure of other generative families — coarse-to-fine token scales, straight transport paths — offers control points for representativeness that diffusion sampling does not.

How the family developed

The papers that moved the family, in the role each one played.

Origin

Path-guided flow matching · arXiv 2026

Presented as the first flow-matching framework for generative distillation: class-conditional transport in a frozen VAE's latent space, with ODE-consistent guidance that lands trajectories on assigned prototypes in few steps.

HIERAMP · CVPR 2026

Uses a visual autoregressive model's coarse-to-fine scales, injecting class tokens at each scale to amplify discriminative structure rather than optimizing global proximity.

What it gets right

  • Coarse-to-fine or path-level control gives a natural place to inject dataset-level guidance.

Where it is weak

  • Very recent; few papers, and evidence is concentrated on ImageNet-style benchmarks.

Applied to

Papers per year

26

Competing answers

All papers in this family 3

Papers not already discussed above, ordered by tier, then newest first.Open in the explorer →

2026-01

ProtoVAR — ProtoVAR: Efficient Dataset Distillation via Prototype-Guided Visual Autoregressive Modeling

Mingyu Wang, Wei Jiang · ICML 2026paper ↗

Replaces diffusion sampling with a visual autoregressive (VAR) model guided by multi-scale class prototypes, adding a fast filtering stage that keeps only the most informative generated samples; reported to run up to 70x faster than diffusion-based generative distillation while remaining competitive on ImageNet-scale benchmarks.