The field in 2026

The short answer. Dataset condensation spent 2018–2023 learning how to make a tiny synthetic set train like a large dataset, and 2023–2025 learning to do it at ImageNet scale. Scale came from taking the learner out of the synthesis loop, first with a frozen teacher and then with a frozen generator. That same move shifted much of the information from the images into a teacher that relabels every crop at evaluation time. By 2026, controlled studies show that at ImageNet-1K scale most of the reported accuracy is set by that teacher and the training recipe. Decoupled, diffusion-based and even random real images converge to within a few points of each other, and well-chosen real subsets often beat synthetic ones under hard labels. What survives is real but narrower: synthesis helps most at small scale, on easy datasets, and when it stays anchored to real data. Multimodal condensation is younger and has a different bottleneck. It has not scaled past about a thousand image–text pairs, and its papers do not share an evaluation protocol.

This essay is the atlas’s point of view, argued from papers read in full. Claims link to the papers they rest on. Numbers carry the protocol that produced them and are never ranked across protocols. Where a statement is judgment rather than a reported finding, it says so.

1. How to think about the field

Every condensation paper answers four questions, and the atlas’s taxonomy keeps them separate. What is condensed (images, image–text pairs, graphs, video…)? How is the synthetic set obtained? What is stored, and what is the label? And what kind of contribution is the paper?

The “how” is best read as a ladder of five paradigms, ordered by how much of the learner sits inside the synthesis loop:

Moving down the ladder, cost fell and scale rose. The central argument of this essay is that the same move relocated where the information lives.

2. Unimodal condensation

2.1 The small-scale era: fidelity versus cost (2018–2023)

Dataset Distillation posed the task by backpropagating through training, and showed that ten synthetic MNIST images could train a network from a fixed initialization. The next five years were a search for surrogates that kept enough of that objective while getting rid of its cost.

• Closed-form inner solvers replaced training with kernel ridge regression. KIP used neural tangent kernels; FRePo used features from a pool of trained networks, matching MTT’s accuracy on CIFAR-100 at one image per class about 100× faster and with roughly 10× less memory. RCIG pushed low-budget accuracy furthest (39.3% on CIFAR-100 at 1 IPC). The advantage of this family shrinks as budgets grow.

• Gradient matching (DC, DSA) matched one training step at a time. It was cheap, but short-sighted, and DCC proved that class-wise gradient matching can fall below random selection when much of the data is task-irrelevant.

• Trajectory matching (MTT) matched long stretches of expert training. Each successor fixed a named weakness:

  • accumulated error (FTD)
  • memory, with an exact constant-memory gradient that reached ImageNet-1K, where soft labels alone added 7–13 points (TESLA)
  • failure at larger budgets, with difficulty-aligned segments that gave the first lossless results at 500+ images per class (DATM)
  • misaligned information (PAD)

  The price was memory and a large up-front investment in expert trajectories. The 2023 TPAMI review records MTT needing 47 GB to unroll 30 steps on CIFAR-10 at 50 IPC (Yu et al.). PAD’s own table shows the most refined trajectory matcher losing to decoupled RDED on CIFAR-100 at 1 IPC and Tiny-ImageNet at 1 and 10 IPC.

The accuracy progression of this era is clearest when label regimes are separated. On CIFAR-10 with a ConvNet, full-data accuracy is 84.8% (each paper’s own table):

With hard labels, the strongest result belongs not to a surrogate but to RaT-BPTT. It returns to the original bi-level objective and stabilizes it with randomly placed truncation windows, which suggests the faithful objective was under-engineered rather than wrong. The best numbers overall add learned labels, which foreshadows the label question that came to dominate the field.

• Distribution matching (DM) dropped the inner loop, matching class-mean features in random networks. It was far cheaper and initially weaker. Its successors added back what the mean leaves free: all moments (M3D), characteristic functions (NCFM), Wasserstein barycenters (WMDD), and style statistics with intra-class diversity (DDM). NCFM reports 71.8% on CIFAR-10 at 10 IPC, ahead of the trajectory-matching results in its table, though that table does not separate the metric from an optional soft-label step. Later members replaced random networks with pretrained ones, quietly moving toward the decoupled paradigm.
• What is stored turned out to be independent of all of this. Parameterizations (IDC, HaBa, FreD, DDiF) work with gradient, distribution and trajectory matching, and their gains hold under explicit equal-storage accounting.

Two findings from this era matter more in hindsight than the leaderboards did. DC-BENCH showed in 2022 that augmentation alone moved results by 6–10 points. With the best augmentation, k-center selection beat gradient matching on CIFAR-10 at 10 IPC. Condensation’s edge over random selection appeared only below about 200 images per class. And none of the synthesis methods preserved architecture rankings for neural architecture search, one of the applications they were promoted for. What is dataset distillation learning? found that bi-level and matching methods store something closer to early training dynamics than a representative sample of the data. Distilled images stop helping, and can even hurt, once they are mixed with real data.

2.2 The scale break: two answers to ImageNet-1K (2023)

In early 2023 the reviews treated ImageNet-1K as barely reachable. TESLA’s 27.9% at 50 IPC was the best entry in the TPAMI review’s table. Two answers appeared within months. TESLA kept trajectory matching and removed its memory bottleneck. SRe2L removed the bi-level loop entirely: train a teacher once, synthesize images against its batch-norm statistics with no student in the loop, and train students on the teacher’s per-crop soft labels. It reported 46.8% at 50 IPC with ResNet-18 and 60.8% with ResNet-101, with synthesis 16–52× faster than MTT. No survey written before it had anticipated abandoning the loop, and the field followed it.

Both answers introduced soft labels into the headline number. In hindsight, that shared choice mattered more than the difference between them.

2.3 Decoupled synthesis: more teacher signal, or more realism? (2024–2026)

After SRe2L the family agreed on a diagnosis: independently recovered images of one class, matched against the same global statistics, look alike. It split on the cure.

• Inversion added teacher signal: augmentation curricula (CDA), several backbones (G-VBSM), performance-weighted committees (CV-DD), architecturally split teachers (PRISM), directed teacher perturbation (DWA), staggered optimization (DELT), audited schedules (EDC) and real-pixel residuals (FADRM).
• Realistic selection stopped inverting: RDED composes images from teacher-scored real patches, NRR-DD refines only non-critical regions, and CIM anchors a feature objective on real images.

With teacher relabeling, the tuned inversion methods now edge past RDED, clustering at 46–50% on ImageNet-1K at 10 IPC with ResNet-18. Without relabeling, the ranking inverts. SRe2L falls to 1.1% and G-VBSM to 0.8%, while RDED keeps 19.7% and CIM 22.0% (CIM, Table 7). CIM’s explanation is the most useful idea the family has produced. Relabeling is reliable only when synthetic images stay close enough to the real distribution that the teacher’s predictions on them still mean something. Realism is not cosmetic: it is what makes the teacher’s labels valid.

2.4 Generative priors: from searching latents to choosing samples (2022–2026)

Generative condensation went through three spending choices.

1. Optimizing GAN latents (IT-GAN, GLaD, H-GLaD) bought cross-architecture generalization. GLaD’s ablation shows the benefit comes from any structured generator manifold: generators trained on unrelated data, or even randomly initialized, keep most of it. It still paid the full matching cost.
2. Storing a generator (DiM, D2M) made redeploying at a new budget nearly free, but stayed at 128×128.
3. Sampling a diffusion model broke the scale ceiling. Minimax fine-tuned for representativeness and diversity, D4M used latent prototypes, and training-free guidance followed (MGD³, D3HR, CaO₂, IGD).

The 2026 papers are mostly critiques of the 2025 ones, which is a sign of a maturing family. ManifoldGD shows that Euclidean mode guidance leaves the data manifold. CoDA argues guidance methods owe their success to a diffusion model already trained on the target dataset, and reaches 60.4% on ImageNet-1K at 50 IPC with an off-the-shelf SDXL instead. Learnability-guided diffusion measures 80–90% overlap between successively generated increments, showing that good samples do not make a good set. IMS3 names the underlying tension: a model trained for likelihood over-represents dense regions and under-covers the class boundaries a classifier needs. Visual autoregressive and flow priors (HIERAMP, path-guided flow matching) bring cheaper sampling and new control points, but have not yet been compared head-to-head with the strongest diffusion methods.

Under matched soft labels, diffusion methods consistently beat decoupled baselines in their own tables. Under matched hard labels at very low budgets, GAN-latent optimization beats them decisively (ImageNette at 1 IPC: 45.4% for H-GLaD against 22.8% for Minimax and 15.2% for D4M).

2.5 The reckoning: what the numbers were measuring (2024–2026)

Four lines of evidence appeared within two years, from different groups using different method sets.

Labels. Random real images given teacher soft labels match state-of-the-art synthetic images (A label is worth a thousand images). Evaluating the same ImageNet-1K subsets at 50 IPC under three regimes gives this (Hard truths about soft labels):

With relabeling, every method and coreset tested falls between 52% and 59%, against 62.5% for the full training set. RDED’s own images score 42.0% with soft labels and 16.3% with one-hot labels (NRR-DD).

Recipe. SRe2L at ImageNet-1K, 10 IPC and ResNet-18 is reported as 21.3%, 31.1%, 41.9% and 43.1% by four papers. Unifying batch size and learning-rate schedule shrinks the published spread across eight methods from 27.3 to 6.7 points (RD³). Changing only the optimizer and loss moves the same images and labels from 1.9% to 53.4% (GIFT).

Baselines. Under each method’s own recipe, several soft-label methods score below random real images (DD-Ranking). On ImageNet-1K with hard labels, a diversity-aware coreset beats all seven recent distilled sets tested, at 10 IPC (18.9% vs 15.6%) and at 50 IPC (43.5% vs 39.3%) (Do distilled sets outperform coresets?). Coresets are also cheaper, and their cost does not grow with the budget.

Storage. Soft labels exceed the images by more than 30× (LPLD), and the generators many methods rely on cost thousands of accelerator-days to train (CoDA). Neither cost is counted in images-per-class.

2.6 Where unimodal condensation ended up

Convergence. At ImageNet-1K with 50 IPC, ResNet-18 students and teacher relabeling, the best decoupled methods report about 58–61%, the best diffusion methods about 60–63%, and random real images with the same relabeling 58.1%. The paradigms have converged on a ceiling set largely by the teacher and the recipe. (This compares figures from different papers, all using teacher relabeling on the same benchmark, so treat it as a band, not a ranking.)

What survives. Synthesis still earns its keep:

• At small scale with hard labels. DATM beats k-center on Tiny-ImageNet, 29.5% vs 22.0%, though the gap disappears with soft labels.
• On small, easy datasets. Diffusion synthesis beats coresets on ImageNette even with hard labels.
• When it stays anchored to real data. RDED and CIM are what survive hard-label evaluation.
• In compressed label regimes. Two scalars per crop recover about 70% of the soft-vs-hard gap (NRR-DD), and a storage-free hard-label phase beats pruned labels at equal storage (HALD).

Selection and synthesis are merging. Nearly every strong 2025–2026 method starts from real content: real-image initialization, patch selection, prototypes or inverted latents. In the atlas’s judgment, the useful question is no longer “distillation or coresets?” but how much synthetic refinement to add on top of a well-chosen real subset, and how that should be evaluated.

For a practitioner (judgment). If the goal is a small ImageNet-scale training set, start from a diversity-aware coreset or a realistic-selection method, evaluate with hard labels as well as soft, and count label storage. Use diffusion synthesis when real data cannot be shared or the dataset is small and easy, and check how much depends on a generator trained on your target domain.

3. Multimodal condensation: a different problem

3.1 What changes

An image–caption corpus has no classes, so the per-class structure that powers image condensation disappears. The target is a joint distribution, the learner is contrastive (a pair’s value depends on the pairs it is contrasted with), and the correspondence between images and captions is itself information that condensation can lose. Text is distilled as continuous embeddings rather than readable sentences. See the vision–language setting for details.

3.2 Four bets

Trajectory matching (2023–2026).

• MTT-VL matched image- and text-encoder trajectories jointly under a contrastive loss.
• LoRS recognized that a caption can describe many images and learned a low-rank similarity matrix.
• RepBlend named modality collapse: under heavy compression, contrastive supervision concentrates each modality’s representations while widening the gap between them. It fixed this while cutting per-iteration time about 6.7×.
• PTM-ST replaced a single unstable trajectory with phase-specific teachers.

Distribution matching (2025–2026) argued that trajectory replay, not optimization, was the cost.

• CovMatch matches image–text cross-covariance without expert trajectories, which makes a trainable text encoder affordable. Its ablation shows freezing BERT alone drops Flickr30K mean recall at 500 pairs from 38.4 to 29.4.
• MDM reports 93–98% less distillation time than LoRS.
• RAHA aligns a low-rank shared subspace in hyperbolic space.

Generative. EDGE builds correspondence into a fine-tuned diffusion model. It is about 20× cheaper than LoRS but does not consistently beat it on recall.

Training-free. PDS clusters CLIP image and text embeddings, matches the prototypes, and decodes with unCLIP, with no optimization at all.

Audio–visual and omnimodal work compressed the same arc into four papers. AVDD used explicit cross-modal losses on random networks; DAVDD moved to pretrained encoders; ImageBindDC condensed inside an already aligned joint space, beating its predecessor on accuracy and speed together; and HoPA reached three modalities at once.

3.3 Why the numbers cannot be pooled

There are at least three evaluation recipes: NFNet with frozen BERT (LoRS, RepBlend, EDGE, PTM-ST), NFNet with trainable BERT (CovMatch, RAHA), and a CLIP text tower (PDS). Even within one recipe, two re-runs of LoRS at 500 pairs disagree (RepBlend reproduces LoRS’s published text→image R@1 of 10.0; PTM-ST reports 12.7), and no paper documents why.

3.4 Where multimodal condensation ended up

Cost has collapsed: expert trajectories no longer look necessary, and generative and training-free methods remove optimization almost entirely. Accuracy and scale have not moved in the same way. No image–text method has been shown beyond roughly a thousand pairs of Flickr30K or COCO, and EDGE’s CC3M result at 1,000 pairs sits near the random floor. Recall stays far below full-data training at every budget; MTT-VL reports a full-data image→text R@1 of 33.9 on Flickr30K against 13.3 for its best distilled set. Multimodal condensation has not had its SRe2L moment. None of the 2024–2026 evaluation studies that reshaped image condensation has yet been run for it.

4. Beyond images

The same ideas were carried to other data, and the transfers were informative.

• Graphs. Gradient matching (GCond) gave way to structure-free trajectory matching (SFGC, GEOM). A striking finding is that synthesized structure is largely optional. GC-Bench finds several leading methods run out of memory at realistic scale, and a 2026 position paper argues the field needs a reset: its methods require full-graph training and its compression metrics hide the cost.
• Video. Naive frame-wise matching barely learns motion; a single distilled still frame recovers most of the achievable accuracy. The fix was to store appearance and motion separately (static–dynamic disentanglement). A large study finds the labeling scheme changes accuracy more than the algorithm (video condensation study).
• Text. Discrete tokens break gradient-based synthesis. Distilling into a generator (DiLM) or constraining optimization to readable tokens (GRADMM) replaced embedding-space methods, whose outputs collapse to near-chance accuracy once decoded into text. A 2025 report names benchmark standardization as the main gap.
• Time series. Forecasting is regression, and classification-style objectives miss a term that measures predicted values (CondTSF); for classification, matching in the frequency domain matters (time-series condensation).
• Pre-training and foundation models. With no labels, matching becomes regression on representations (KRR-ST). Distilling data for a linear probe on a frozen foundation model reaches within about 8 points of full ImageNet-1K with one image per class (linear gradient matching).
• Detection. Images-per-class is meaningless when an image holds many objects. Inversion (DCOD) was overtaken by training-free copy-paste-and-filter (OD3).

Across these settings, gradient and trajectory matching transferred cleanly, and the label and evaluation confounds reappeared independently in video, graphs and text. No generative-prior method yet exists for graphs, video, time series or detection.

5. Cross-cutting lessons

1. The coupling ladder is also a map of where information moved. Each paradigm removed part of the learner from synthesis and moved information into something needed at evaluation time: expert trajectories, a relabeling teacher, a pretrained generator. Scale was bought partly by relocating information, not only by compressing it.
2. Realism is functional. It began as a side benefit (GLaD, RDED) and became a requirement: relabeling is reliable only near the real distribution, inverted images collapse without soft labels, and methods anchored in real content survive hard-label evaluation.
3. The set is the unit, not the sample. Diversity failures were rediscovered in every paradigm: mean matching, look-alike recovered images, redundant diffusion samples, and trajectory matching at high budgets. Methods that treat samples independently keep hitting the same wall.
4. Budgets are counted in the wrong unit. Images-per-class leaves out soft labels, decoders, generators and teachers, along with their pre-training. Synthesis cost grows with the budget; selection cost does not.
5. The reviews predicted scale; the field got a reckoning. Every survey predicted more scale, more modalities and better theory (Liu & Du, Lei & Tao, Sachdeva & McAuley). What dominated 2025–2026 was evidence about how much of the previous gains lived in labels and recipes.
6. Trust has not caught up. Condensation’s privacy benefit is disputed (Privacy for free versus No free lunch); soft labels can leak what a teacher memorized (soft-label leakage); and condensation tends to worsen fairness toward minority groups (FairDD). See Open problems.

6. Open bets for 2026–2027

These are the atlas’s judgments, each tied to the evidence above, and will be revisited.

• Hard-label, compute-matched evaluation becomes standard, and reorders the leaderboard. The tools exist (DD-Ranking, RD³); real-anchored methods and strong coresets are best placed.
• Condensation and selection become one design space: a well-chosen real subset plus a small synthetic or refined residual, as in CIM, NRR-DD and pushforward quantization.
• Generative condensation moves to set-level objectives and starts pricing in the pre-training of its generators.
• Labels become explicit, compact and audited artifacts, counted in the budget and checked for leakage.
• Multimodal condensation gets a shared benchmark before it gets a leaderboard, and a decoupled or generative method finally pushes it beyond Flickr30K and COCO scale. This is the field’s clearest open opportunity.

How this essay was written

The essay rests on full-text reading of the landmark, core and most-cited notable papers in every family and setting. Structured digests of each group are in the repository under docs/digests/. Seven surveys and three 2025–2026 position papers and reports were also read. Papers whose full text could not be obtained (a few OpenReview-only 2026 papers) are not used as evidence. Corrections are welcome through the repository’s issue tracker.