Dataset Condensation Atlas

Settings · multimodal

Image–text pairs (vision–language)

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.

One synthetic sample

An image paired with one or more captions (often stored as text embeddings), sometimes accompanied by a learned image–text similarity structure instead of strict one-to-one pairing.

Standard evaluation

Train or fine-tune an image encoder and a text encoder or projection contrastively on the synthetic pairs, then report image→text and text→image Recall@K on the Flickr30K or MS-COCO test split.

Budget

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

Benchmarks

Flickr30K · MS-COCO

What changes when there are no classes

Class-conditional condensation gets structure for free: statistics, gradients and trajectories are matched class by class, and budgets are counted per class. An image–caption corpus has no classes. The distilled set has to approximate a joint image–text distribution, and the model trained on it is contrastive, so the value of a pair depends on the other pairs it is contrasted with. Three consequences shape the whole setting:

Four bets, and what each generation fixed

Trajectory matching. MTT-VL opened the setting in 2023 by matching expert trajectories of an image encoder and a text encoder jointly under a contrastive loss, with hard one-to-one pairs. Each successor fixed the previous failure:

Distribution matching. The next group argued that the expensive part was trajectory replay itself, not optimization.

Generative synthesis. EDGE builds correspondence into the generator. It fine-tunes Stable Diffusion with contrastive and diversity losses, then samples pairs. Its total pipeline is about 20× cheaper than LoRS on COCO (7.7 vs 151.6 GPU-hours at 500 pairs), but it does not consistently beat LoRS on recall (Flickr30K, 500 pairs: image→text R@1 13.3 vs 15.5; text→image R@1 6.7 vs 10.0).

Training-free synthesis. PDS optimizes nothing. It clusters CLIP image and text embeddings, matches the two sets of prototypes, and decodes images with unCLIP, at about 10 seconds per synthesized image.

Where each method establishes correspondence

WhereMethods
In the distillation objectiveMTT-VL (contrastive loss inside trajectory matching), MDM (InfoNCE plus hypersphere matching)
In a stored similarity structureLoRS, RepBlend, PTM-ST
In the feature geometry that is matchedCovMatch (cross-covariance), RAHA (shared low-rank subspace, hyperbolic)
In a fine-tuned generatorEDGE
After the fact, by matching per-modality prototypesPDS

Reading results: three incompatible protocols

The papers fall into at least three evaluation recipes, and numbers should only be compared within one.

ProtocolImage encoder at evaluationText encoderPapers
NFNet, frozen BERT with a trained projectionpretrained, trainedBERT-base, frozenLoRS, RepBlend, EDGE, PTM-ST
NFNet, trainable BERTpretrained, trainedBERT-base, trainedCovMatch, RAHA
CLIP text tower with other image backbonesResNet-50 / ViT-Ti, trainedCLIP, frozenPDS

Within the frozen-BERT protocol on Flickr30K at 500 pairs, text→image R@1 is 10.0 for LoRS, 17.0 for RepBlend and 16.0 for PTM-ST (each paper’s own table). Even this is imperfect: RepBlend’s re-run of LoRS reproduces LoRS’s published numbers exactly, while PTM-ST’s re-run gives different ones (12.7 instead of 10.0), and no paper documents why. PDS compares against LoRS and TESLA-VL re-implemented with a CLIP text tower, so its table is internally consistent but cannot be placed next to the others.

Where the setting stands

The open problems on incompatible protocols and scale follow from this. The other meaning of “multimodal”, where a vision–language model is a tool for condensing ordinary image datasets (VLCP, EVLF, CoDA), is covered under image classification.

What is different here

  • There are no classes. The target is the joint image–text distribution, so class-wise matching has no direct analogue, and the correspondence between the two modalities is itself something the distilled set has to preserve.
  • The downstream learner is contrastive, so the value of one pair depends on the pairs it is contrasted against. Distilling pairwise similarity structure rather than assuming one caption per image is one response to this.
  • Methods split into optimization-based ones (trajectory, distribution, covariance or analytic parameter matching through the encoders) and generative ones (a pretrained diffusion prior synthesizes the images, conditioned on real, prototypical or generated text).
  • Text is usually distilled in embedding space, so the released artifact is not always a readable dataset.

Comparison pitfalls

  • Evaluation protocols are not shared across papers: pretrained versus frozen image backbones, frozen BERT versus CLIP text encoders, trainable versus fixed text towers. Numbers from different protocols are not comparable, and the difference for a single method can be large.
  • Extra stored structure (similarity matrices, soft labels) is rarely counted against the pair budget.
  • Recall at 100–1000 pairs remains far below full-data recall; gains over weak coreset baselines can overstate practical usefulness.

Start with

MTT-VL · LoRS · RepBlend · CovMatch · EDGE · PDS · MDM · RAHA

Methods by family 12

A paper appears under every family it belongs to. Oldest first within a family.

Bi-level performance matching

Meta-learning through unrolled training 1

2026-07

InfMatch — Dataset Distillation by Influence Matching

Haoru Tan, Wang Wang, Sitong Wu et al. · CVPR 2026Vision–languagepaper ↗code ↗

Replaces process-level surrogates (per-step gradients or training trajectories) with a differentiable, sample-level estimator of each synthetic point's influence on the converged model parameters, computed in linear time by unrolling the optimization dynamics with a first-order Taylor approximation, then learns the synthetic set by matching this influence to that of the real data; reports 31.5% accuracy on Tiny-ImageNet at IPC-10 (a 4.7-point gain over NCFM) and extends to vision-language distillation on Flickr30K.

Bi-level performance matching

Kernel and closed-form inner solvers 1

2026-01

MMDD — Efficient Multi-modal Dataset Distillation via Analytic Parameter Matching

Deyu Bo, Xinchao Wang · ICML 2026notableVision–languagepaper ↗code ↗

MMDD replaces the bi-trajectory distillation used by prior multimodal dataset distillation methods (aligning expert and student trajectories per modality, which needs many stored checkpoints and double backpropagation) with analytic parameter matching, directly matching the closed-form analytic parameters of the modal projectors, which the paper connects theoretically to matrix whitening. It caches only two matrices instead of trajectory checkpoints and computes the match in a single forward pass, reporting up to 65x storage reduction and 9.6x distillation speedup while scaling to 1000 synthetic pairs on Flickr30k and MS-COCO.

Surrogate matching

Trajectory matching 5

2023-08

MTT-VL — Vision-Language Dataset Distillation

Xindi Wu, Byron Zhang, Zhiwei Deng et al. · TMLR 2024landmarkVision–languagepaper ↗code ↗

This is the first vision-language dataset distillation method, extending trajectory matching to image-text pairs by jointly distilling them in a contrastive formulation (since there are no discrete classes to condition on) and using LoRA matching for efficient trajectory matching in large vision-language models. Against adapted vision-language coreset-selection baselines, the paper reports nearly doubling Flickr30K image-to-text recall@1 (5.6% to 9.9%) while using 100 distilled pairs versus 1000 selected ones.

2024-06

LoRS — Low-Rank Similarity Mining for Multimodal Dataset Distillation

Yue Xu, Zhilin Lin, Yusong Qiu et al. · ICML 2024coreVision–languagepaper ↗code ↗

LoRS distills a ground-truth image-text similarity matrix alongside the synthetic pairs, instead of assuming one caption strictly matches one image, and stores that similarity structure efficiently via low-rank factorization. The paper reports significant improvements over prior vision-language distillation algorithms and proposes LoRS as a foundational synthetic-data setup for image-text dataset distillation.

2025-05

RepBlend — Beyond Modality Collapse: Representations Blending for Multimodal Dataset Distillation

Xin Zhang, Ziruo Zhang, Jiawei Du et al. · NeurIPS 2025notableVision–languagepaper ↗

RepBlend identifies "modality collapse" in multimodal dataset distillation, over-concentrated intra-modal representations and an enlarged cross-modal gap, as a conflict between dataset distillation's over-compression and contrastive cross-modal supervision, and addresses it by weakening overdominant cross-modal supervision through representation blending to restore intra-modal diversity, plus symmetric projection trajectory matching that synchronizes optimization dynamics across modality-specific projection heads to correct asymmetric supervision. On Flickr30K and MS-COCO the paper reports consistent gains over prior multimodal distillation methods (e.g. +9.4 IR@10, +6.3 TR@10 at the 100-pair setting) with up to 6.7x faster distillation.

2026-01

AMD — Asynchronous Matching with Dynamic Sampling for Multimodal Dataset Distillation

Ding Qi, Jian Li, Shuguang Dou et al. · ICLR 2026notableVision–languagepaper ↗

Proposes AMD for image-text dataset distillation, performing asynchronous trajectory matching that decouples the starting points of image and text expert trajectories to accommodate their different learning speeds, and replacing random prototype initialization with a Semantics-Aware Prototype Mining module that clusters the feature space into representative prototypes; reported Recall@1/5/10 gains are 4.5, 9.6 and 10.9 points on Flickr30K at 200 distilled pairs.

2026-03

PTM-ST — Multimodal Dataset Distillation via Phased Teacher Models

Shengbin Guo, Hang Zhao, Senqiao Yang et al. · ICLR 2026notableVision–languagepaper ↗code ↗

PTM-ST distills vision-language pairs by matching a stage-aware model of the teacher's training dynamics rather than a single trajectory, using a shortcut-based trajectory construction strategy to fit distinct training phases and stabilize the match, addressing performance gaps between phases and unstable teacher trajectories seen in prior multimodal trajectory-matching methods. On Flickr30K and MS-COCO the paper reports surpassing prior state-of-the-art methods by up to 13.5 percentage points absolute, with an average gain of 9.53 points on Flickr30K, while also reducing storage overhead.

Surrogate matching

Distribution and feature matching 3

2025-10

CovMatch — CovMatch: Cross-Covariance Guided Multimodal Dataset Distillation with Trainable Text Encoder

Yongmin Lee, Hye Won Chung · NeurIPS 2025notableVision–languagepaper ↗code ↗

Proposes CovMatch for image-text dataset distillation, which fixes the encoders at each distillation step and matches the cross-covariance between real and synthetic image and text feature matrices, together with per-modality feature-distribution regularization, instead of unrolling trajectories; unlike prior methods that freeze the text tower, CovMatch keeps a BERT-base text encoder trainable end to end (only its embedding layer is frozen) alongside a pretrained NFNet image encoder. On Flickr30K at 500 pairs it reports 38.4% mean recall (Table 2), a 6.8-point absolute gain over the LoRS baseline (31.6%), and 19.6% mean recall on COCO at 500 pairs, a 6.1-point gain over LoRS.

2026-05

MDM — Multimodal Distribution Matching for Vision-Language Dataset Distillation

Jongoh Jeong, Hoyong Kwon, Minseok Kim et al. · CVPR 2026notableVision–languagepaper ↗code ↗

MDM distills vision-language pairs by initializing synthetic image-text pairs from clusters in the joint embedding space, forming a mixed teacher by interpolating independently fine-tuned models in weight space, and matching real and synthetic joint distributions on the unit hypersphere with a geometry-aware objective that exploits cross-modal agreement and discrepancy directions alongside symmetric contrastive learning. The paper reports compact synthetic sets that preserve multimodal semantics and remain robust under cross-architecture evaluation on image-text retrieval benchmarks.

2026-06

RAHA — Rank-Aware Hyperbolic Alignment for Vision-Language Dataset Distillation

Jongoh Jeong, Sun-Kyung Lee, Kuk-Jin Yoon · ECCV 2026notableVision–languagepaper ↗code ↗

RAHA argues that full-dimensional Euclidean alignment is too restrictive for the rank-deficient image-text correlation typical of vision-language distillation (shared semantics concentrate in a low-dimensional range, with the remainder in a weakly correlated residual subspace), and instead lifts representations to hyperbolic space, optimizing distilled pairs with asymmetric objectives that enforce geodesic alignment on the shared range while regularizing the residual subspace to keep modality-private diversity. The paper reports competitive cross-modal retrieval and improved transfer robustness under fixed budgets versus prior similarity-level low-rank alignment (LoRS).

Generative priors

Diffusion-based synthesis 2

2025-09

EDGE — Efficient Multimodal Dataset Distillation via Generative Models

Zhenghao Zhao, Haoxuan Wang, Junyi Wu et al. · NeurIPS 2025coreVision–languagepaper ↗code ↗

Proposes EDGE, which fine-tunes a pretrained Stable Diffusion v1.5 model for image-text dataset distillation by replacing its denoising loss with a bidirectional contrastive loss that aligns generated image and text embeddings and a minimax diversity loss that pushes different pairs' embeddings apart, then generates synthetic pairs by sampling from the fine-tuned model; evaluated with a pretrained NFNet image encoder and a frozen BERT-base text encoder, it reports on Flickr30K at 500 pairs 21.0/30.5% image-retrieval Recall@5/10 and 35.6/47.5% text-retrieval Recall@5/10 (Table 1), running about 18x faster than the trajectory-matching baseline LoRS.

2026-02

PDS — Multimodal Dataset Distillation Made Simple by Prototype-Guided Data Synthesis

Junhyeok Choi, Sangwoo Mo, Minwoo Chae · ICLR 2026coreVision–languagepaper ↗code ↗

PDS is a learning-free multimodal distillation framework that uses CLIP to extract aligned image-text embeddings, derives prototypes from them, and synthesizes images with an unCLIP decoder, avoiding the full-dataset training and joint pixel/text optimization that architecture-dependent distillation methods require. The paper reports consistently outperforming optimization-based dataset distillation and subset-selection baselines and achieving state-of-the-art cross-architecture generalization.

Orthogonal design choices

Label distillation and soft labels 1

2024-06

LoRS — Low-Rank Similarity Mining for Multimodal Dataset Distillation

Yue Xu, Zhilin Lin, Yusong Qiu et al. · ICML 2024coreVision–languagepaper ↗code ↗

LoRS distills a ground-truth image-text similarity matrix alongside the synthetic pairs, instead of assuming one caption strictly matches one image, and stores that similarity structure efficiently via low-rank factorization. The paper reports significant improvements over prior vision-language distillation algorithms and proposes LoRS as a foundational synthetic-data setup for image-text dataset distillation.

Orthogonal design choices

Optimization and training recipes 2

2026-03

PTM-ST — Multimodal Dataset Distillation via Phased Teacher Models

Shengbin Guo, Hang Zhao, Senqiao Yang et al. · ICLR 2026notableVision–languagepaper ↗code ↗

PTM-ST distills vision-language pairs by matching a stage-aware model of the teacher's training dynamics rather than a single trajectory, using a shortcut-based trajectory construction strategy to fit distinct training phases and stabilize the match, addressing performance gaps between phases and unstable teacher trajectories seen in prior multimodal trajectory-matching methods. On Flickr30K and MS-COCO the paper reports surpassing prior state-of-the-art methods by up to 13.5 percentage points absolute, with an average gain of 9.53 points on Flickr30K, while also reducing storage overhead.

2026-07

InfMatch — Dataset Distillation by Influence Matching

Haoru Tan, Wang Wang, Sitong Wu et al. · CVPR 2026Vision–languagepaper ↗code ↗

Replaces process-level surrogates (per-step gradients or training trajectories) with a differentiable, sample-level estimator of each synthetic point's influence on the converged model parameters, computed in linear time by unrolling the optimization dynamics with a first-order Taylor approximation, then learns the synthetic set by matching this influence to that of the real data; reports 31.5% accuracy on Tiny-ImageNet at IPC-10 (a 4.7-point gain over NCFM) and extends to vision-language distillation on Flickr30K.

Analyses, benchmarks, surveys and applications

2025-01

Shang et al. survey — Dataset Distillation in the Era of Large-Scale Data: Methods, Analysis, and Future Directions

Xinyi Shang · arXiv 2025Vision–languagepaper ↗

This survey identifies four shifts since 2023 that earlier surveys predate: scaling distillation to large datasets, moving from optimizing distilled samples alone to improving the whole distillation pipeline, developing theoretical and empirical understanding of core mechanisms, and expanding from images to diverse modalities. It analyzes methods for each stage, benchmarks contemporary methods on large-scale datasets, summarizes distillation across unimodal and multimodal data, and lays out open challenges and directions.