Settings

The canonical setting. Condense a labeled image dataset so that a freshly initialized classifier trained only on the synthetic set approaches the accuracy of one trained on the full data.

Budget: images per class (IPC)

Condense a paired image–caption corpus so that a dual encoder trained on the synthetic pairs retrieves nearly as well as one trained on the full corpus.

Budget: number of image–text pairs (typically 100–1000)

Condense datasets whose samples span audio, vision and sometimes further modalities, while preserving the correspondence between them.

Budget: samples per class

Condense video datasets for action recognition or temporal understanding, where most of the cost and most of the redundancy lie in the time axis.

Budget: clips per class

Graph condensation: synthesize a small graph (or a few small graphs) on which a graph neural network trains to the accuracy it reaches on the original large graph.

Budget: condensation ratio (fraction of nodes kept)

Condense text datasets for classifiers or language-model fine-tuning, where the discreteness of tokens breaks the gradient-based synthesis that works for pixels.

Budget: examples per class

Condense sequential datasets for forecasting or classification, where the target is often a regression output rather than a class label.

Budget: number of synthetic sequences or windows

Point clouds, tabular data, recommender interactions, speech, EEG, event streams, reinforcement learning and more, each with a handful of papers adapting condensation to its structure.

Budget: domain-specific

Condense unlabeled data for self-supervised pre-training, or condense data for training on top of frozen pretrained models, and judge the result by transfer rather than by in-distribution accuracy.

Budget: number of images

Condense datasets whose labels are spatial (boxes, masks, target images), where one image holds many objects and synthesis must preserve layout.

Budget: number of images, or a fraction of the training set