P4 · Generative priors · since 2024 · ascendant

Diffusion-based synthesis

Generate the synthetic set with a pretrained (optionally fine-tuned) diffusion model, steering sampling toward samples that are both representative of the dataset and diverse.

The bet every member shares

A large pretrained diffusion model already captures the data distribution; the condensation problem reduces to choosing where in that distribution to spend a small sample budget.

Defining objective

\hat{\epsilon}(x_t,t)=\epsilon_{\theta}(x_t,t,c_k)\;-\;\lambda_t\,\nabla_{x_t}\,\mathcal{G}\bigl(\hat{x}_0(x_t)\bigr),\qquad \mathcal{S}=\bigl\{x_0^{(k)}\bigr\}_{k=1}^{K}

A generic guided sampler. Members differ in the condition c_k (a class label, a latent prototype or mode, a text prototype, an image–text mode), in the guidance G (representativeness, diversity, influence, learnability, manifold or set-level terms), and in whether the diffusion model is fine-tuned for the purpose.

From optimizing pixels to choosing samples

Every earlier paradigm treats the synthetic set as a variable to optimize. A pretrained diffusion model changes the question. The model can already produce realistic, high-resolution images of each class, so condensation becomes a matter of choosing where in its distribution to spend a small budget, and of steering sampling there. Minimax Diffusion and D4M brought this view to ImageNet-scale benchmarks in 2024. By 2026 it was the family with the most activity, and the most visible arguments.

It inherits from two earlier generative ideas, each of which fixed one problem and kept another:

Optimizing GAN latents (IT-GAN, GLaD) fixed cross-architecture generalization. GLaD’s own ablation shows the benefit comes from constraining synthesis to a structured generator manifold at all; generators trained on unrelated data, or even randomly initialized, keep most of it. But it still paid the full cost of a matching objective through the generator.
Storing a generator (DiM, D2M) made redeploying at a new budget nearly free, but stayed at 128×128 and trained its generator on the target data.

Diffusion sampling removed the scale ceiling. It gave up optimizing each sample for training utility, and much of the family’s later history is about recovering that without paying for it.

How the family developed: each fix exposed the next problem

Fine-tune or pick prototypes (2024). Minimax fine-tunes a DiT with representativeness and diversity terms. D4M clusters latent prototypes of a Stable Diffusion model.
Guide without training (2025). MGD³ discovers modes and guides sampling toward distinct ones, with no fine-tuning. D3HR inverts the dataset into the Gaussian domain with DDIM and samples representative latents there. CaO₂ fixes mismatches between generation and the evaluation objective. IGD guides by the influence a sample would have on training.
The guidance itself is questioned (2026). ManifoldGD shows that Euclidean mode guidance pushes samples off the data manifold, and projects guidance onto its local tangent space. CoDA transplants guidance methods between backbones and argues their success rests on a diffusion model already trained on the target dataset. It then gets comparable ImageNet-1K results from an off-the-shelf SDXL model steered toward a density-discovered “core” of each class. DAP reads representativeness directly from the frozen backbone’s own features.
From samples to sets (2026). Even good per-sample guidance produces redundant sets: learnability-guided diffusion measures 80–90% overlap between successively generated increments and builds the set as a curriculum instead. DMGD adds optimal-transport matching over the whole set. IMS3 names the underlying tension: a model trained for likelihood over-represents dense regions and under-covers the class boundaries a classifier needs.
New generators. HIERAMP uses a visual autoregressive model’s coarse-to-fine scales as separate control points. Path-guided flow matching samples in 32 steps rather than 50 and reports 78% mode coverage against MGD³’s 30%. Neither has yet been compared head-to-head with the strongest training-free diffusion methods.

What the numbers support

Under a matched soft-label protocol, the evidence is consistent. D4M, IGD, D3HR, MGD³ and CaO₂ each beat RDED and SRe2L on ImageNet-1K in their own tables. The 2026 methods cluster tightly. At 50 IPC with ResNet-18 students and teacher relabeling, pushforward quantization and DAP report 62.7%, EVLF on MGD³ 61.9%, DMGD 61.4%, VLCP 60.5% and CoDA 60.4%. That band also contains the best decoupled methods, and it sits a few points above random real images trained with the same kind of relabeling (58.1% in Hard truths about soft labels). At this scale, with a teacher relabeling every crop, the choice between paradigms moves accuracy by a few points at most.

Under hard labels the picture changes. H-GLaD strips soft labels from both sides at very low budgets and finds GAN-latent optimization far ahead of Minimax and D4M (ImageNette, 1 IPC: 45.4% vs 22.8% vs 15.2%). On ImageNet-1K with hard labels, a diversity-aware coreset beats every diffusion method tested (Do distilled sets outperform coresets?). The reversal is ImageNette, where diffusion synthesis beats coresets even with hard labels. Diffusion-based condensation clearly produces better sets than naive generation. Whether its images beat carefully selected real ones at scale is not yet shown.

Open problems specific to this family

Whose prior is it? Most strong results still use a generator pretrained on the target domain. CoDA shows one way around this for ImageNet, but no method has yet removed fine-tuning, target-domain pre-training and per-sample redundancy at once, with a head-to-head comparison.
Likelihood is not utility. Sampling concentrates where the density is high; classifiers need coverage near boundaries. Set-level objectives are the current answer, and they reintroduce training loops.
Cost accounting. Sampling is cheap, but a diffusion model trained on ImageNet cost thousands of accelerator-days, and synthesis time grows with the budget, unlike selection.
Evaluation. Nearly every ImageNet-1K headline uses teacher relabeling. Hard-label numbers are rare and, where they exist, less favorable. See Evaluation.

Branches

Prototypes and modes

Find dense regions (clusters, modes) of the data in a latent or embedding space and generate one sample per region.

D4M, MGD³, CoDA

Fine-tuned diffusion priors

Fine-tune the diffusion model with representativeness and diversity objectives, or with cross-modal contrastive objectives.

Minimax Diffusion, IMS3

Training-free guidance

Leave the model untouched and add guidance terms at sampling time toward informative, learnable or diverse samples.

ManifoldGD, MGD³

Inversion and latent statistics

Invert real images into the diffusion latent space and select or match latents by their statistics.

D3HR, CaO2

Vision–language conditioning

Use text prototypes, captions or text-to-image models to condition generation with richer semantics than class names.

VLCP, CoDA

How the family developed

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

Origin

Minimax Diffusion · CVPR 2024

Fine-tunes a diffusion model with minimax criteria for representativeness and diversity; reports better accuracy at 100 IPC on ImageWoof with under a twentieth of earlier methods' distillation time.

D4M · CVPR 2024

Builds the distilled set from class prototypes in a latent diffusion model's space, independent of any matching architecture, to target cross-architecture generalization.

Improvement

D3HR · ICML 2025

Diagnoses inaccurate distribution matching, noise-induced deviation and separate sampling in earlier diffusion methods; maps dataset latents to a Gaussian domain by DDIM inversion and samples representative latents there.

MGD³ · ICML 2025

Drops fine-tuning with distillation losses: discovers modes, guides sampling toward distinct modes to raise intra-class diversity, and stops guidance early to avoid artifacts.

CaO2 · ICCV 2025

Names objective inconsistency and condition inconsistency between diffusion-based distillation and evaluation; selects samples by probability, then refines their latents for conditional likelihood.

ManifoldGD · CVPR 2026

Training-free guidance that projects the mode-alignment direction onto a local tangent space of the latent manifold at every denoising step, using hierarchical clustering of VAE latents.

Extension

VLCP · ICCV 2025

Adds text prototypes, built from descriptions generated by an open-source language model, to image prototypes when conditioning synthesis.

Challenge

CoDA · ICLR 2026

Points out that most diffusion-based methods need a diffusion model trained on the full target dataset, and instead steers an off-the-shelf text-to-image model toward a density-discovered core distribution; reports 60.4% on ImageNet-1K at 50 IPC.

IMS3 · CVPR 2026

Argues that diffusion models serve likelihood, not discriminative utility, and over-concentrate in high-density regions; inversion-guided fine-tuning widens coverage and subgroup sampling improves class separability.

What it gets right

Cheap relative to matching-based synthesis, often training-free, and it produces realistic high-resolution images.
Scales to ImageNet-1K and is flexible in the budget, since samples are generated, not optimized.

Where it is weak

Samples are not optimized for the learner, so the evaluation usually leans on soft labels from a pretrained teacher as well.
Quality is bounded by the pretrained model's domain; datasets outside it suffer.
Mode-seeking sampling can reduce within-class diversity, which several members exist to correct.

Applied to

Papers per year

Competing answers

All papers in this family 36

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

2026-02

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

Junhyeok Choi, Sangwoo Mo, Minwoo Chae · ICLR 2026coreVision–language paper ↗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.

2025-09

EDGE — Efficient Multimodal Dataset Distillation via Generative Models

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

2026-05

DMGD — DMGD: Train-Free Dataset Distillation with Semantic-Distribution Matching in Diffusion Models

Qichao Wang, Yunhong Lu, Hengyuan Cao et al. · CVPR 2026notablepaper ↗

Proposes training-free Dual Matching Guided Diffusion: Semantic Matching via conditional-likelihood optimization removes the need for an auxiliary classifier, a dynamic guidance mechanism balances diversity against semantic alignment, and an optimal-transport-based Distribution Matching term aligns generated samples with the target distribution's structure, with two efficiency strategies (Distribution Approximate Matching, Greedy Progressive Matching); reports average accuracy gains of 2.1%, 5.4% and 2.4% on ImageNet-Woof, -Nette and -1K over fine-tuning-based diffusion methods.

2026-04

Learnability-guided diffusion — Learnability-Guided Diffusion for Dataset Distillation

Jeffrey A. Chan-Santiago, Mubarak Shah · CVPR 2026notablepaper ↗

Finds that prior diffusion-based distillation methods produce redundant samples -- disjoint subsets of a distilled set share 80-90% overlapping signal -- because they optimize visual diversity or average training dynamics without accounting for sample-to-sample similarity; Learnability-Guided Diffusion instead builds the synthetic set incrementally, training a model on what exists so far and generating new samples guided by learnability scores balanced against validity under a reference model, cutting redundancy by 39.1% and reaching 60.1% on ImageNet-1K, 87.2% on ImageNette and 72.9% on ImageWoof.

2026-03

EVLF — EVLF: Early Vision-Language Fusion for Generative Dataset Distillation

Wenqi Cai, Yawen Zou, Guang Li et al. · CVPR 2026notablepaper ↗code ↗

Observes that late-stage cross-attention fusion in diffusion-based distillation lets text prompts dominate generation, over-correcting samples toward prompt patterns at the expense of visual detail; Early Vision-Language Fusion instead aligns text and visual embeddings with a lightweight cross-attention module at the encoder-to-backbone transition, and plugs into any diffusion-based distillation pipeline with an encoder regardless of denoiser architecture or sampling schedule.

2025-11

Optimizing Distributional Geometry Alignment with Optimal Transport for Generative Dataset Distillation

Xiao Cui, Yulei Qin, Wengang Zhou et al. · NeurIPS 2025notablepaper ↗

Reformulates generative dataset distillation as Optimal Transport distance minimization to capture instance-level and intra-class geometry that global mean/variance matching misses, with three components: OT-guided diffusion sampling that aligns latent distributions, label-image-aligned soft relabeling that adapts label distributions to the complexity of the distilled images, and OT-based logit matching between student outputs and soft-label distributions; reports at least a 4% accuracy gain over the state of the art at IPC=10 on ImageNet-1K across architectures.

2025-10

DAP — Diffusion Models as Dataset Distillation Priors

Duo Su, Huyu Wu, Huanran Chen et al. · ICLR 2026notablepaper ↗

Formalizes the representativeness already implicit in a pretrained diffusion model by quantifying feature-space similarity between synthetic and real data with a Mercer kernel, then uses it as training-free guidance for the reverse diffusion process instead of adding external constraints; reports state-of-the-art fidelity and cross-architecture generalization on ImageNet-1K and its subsets.

2025-01

Pushforward quantization — Dataset Distillation as Pushforward Optimal Quantization

Hong Ye Tan, Emma Slade · ICLR 2026notablepaper ↗

Shows that decoupled/disentangled dataset-distillation methods with an encoder-decoder structure are instances of classical optimal quantization and Wasserstein-barycenter problems, and uses this connection to propose clustering in a diffusion model's latent space to select prototype points, reporting better performance and cross-model generalization than D4M on ImageNet-1K, plus state-of-the-art results using the distilled noise initializations with a stronger diffusion transformer.

2025-01

UniDD — Towards Universal Dataset Distillation via Task-Driven Diffusion

Ding Qi, Jian Li, Junyao Gao et al. · CVPR 2025notableDense prediction paper ↗

UniDD extends dataset distillation beyond classification to detection and segmentation with a task-driven diffusion model, first mining task-relevant information by training task-specific proxy models (Universal Task Knowledge Mining), then using those proxies to guide a diffusion process that generates task-specific synthetic images with localized, region-level control rather than the global, single-class updates of prior generation methods (Universal Task-Driven Diffusion). The paper reports consistently outperforming state-of-the-art methods on ImageNet-1K, Pascal VOC and MS COCO across classification, detection and segmentation.

2025-01

IGD — Influence-Guided Diffusion for Dataset Distillation

Mingyang Chen, Jiawei Du, Bo Huang et al. · ICLR 2025notablepaper ↗code ↗

Frames dataset distillation as controlled diffusion generation guided by a trajectory influence function that links generated samples to downstream training effectiveness, requiring no diffusion retraining; reports 60.3% top-1 at IPC=50 on ImageNet-1K, state-of-the-art among diffusion-based methods.

2026-08

SRG — Self-Supervised Representation-Guided Generative Dataset Distillation

Mingzhuo Li, Guang Li, Linfeng Ye et al. · arXiv 2026Pre-training & transfer paper ↗

Targets distillation for downstream training with frozen pretrained self-supervised encoders and lightweight adapters rather than randomly initialized networks: builds class-wise prototypes from real-image SSL representations and translates three SSL-space objectives (prototype alignment, inter-class discrimination, intra-class assignment) into diffusion guidance, anchoring early denoising to the latent of the nearest real image and later denoising to the SSL objectives. Reports consistent gains over evaluated generative baselines across datasets and IPC settings, with transfer across different pretrained representation spaces.

2026-07

Adaptive Latent Trajectory Anchoring for Action Segmentation Dataset Condensation

Artheme Gauthier-Villar, Guodong Ding, Angela Yao · ECCV 2026Video paper ↗

Condenses action-segmentation video datasets by replacing VAE-based iterative latent optimization with deterministic DDIM latent trajectories anchored by sparse latent points on the noise manifold, and adaptively reallocates the anchoring budget across segments by per-segment reconstruction difficulty; reports performance parity with full-data training at a 2.4% condensation ratio on the Breakfast dataset.

2026-05

D3S2 — D3S2: Diffusion-Guided Dataset Distillation for Semantic Segmentation

Wenjie Zheng, Haoji Hu, Jiali Lu et al. · arXiv 2026Dense prediction paper ↗

Proposes D3S2 for dense-prediction dataset distillation in semantic segmentation: a class-balanced set of masks is selected by a greedy strategy that prioritizes underrepresented classes, then a pretrained layout-to-image diffusion model generates images conditioned on those masks, guided at sampling time by a segmentation-consistency loss and a class-wise feature-matching loss for pixel-level alignment; at a 1% compression rate with Mask2Former (Swin-S) it reports 24.99% mIoU on ADE20K and 35.49% on COCO-Stuff, improving over random selection by 9.34 and 5.70 points respectively.

2026-05

DIVER — DIVER:Diving Deeper into Distilled Data via Expressive Semantic Recovery

Qianxin Xia, Zhiyong Shu, Wenbo Jiang et al. · ICML 2026paper ↗code ↗

Takes distilled images from any prior single-stage method and 'dives deeper' with a pretrained diffusion model in three steps -- semantic inheritance projects the abstract distilled images' high-level semantics into the diffusion latent space to filter out architecture-specific noise, semantic guidance directs the reverse process to preserve that semantics, and semantic fusion applies the guidance only during the concrete phase to avoid artifacts -- improving cross-architecture generalization with runtime comparable to a raw DiT pass on ImageNet at 4GB GPU memory.

2026-01

DPD — Towards Realistic Remote Sensing Dataset Distillation with Discriminative Prototype-guided Diffusion

Yonghao Xu, Pedram Ghamisi, Qihao Weng · arXiv 2026paper ↗code ↗

Introduces dataset distillation to remote sensing image interpretation for the first time with discriminative prototype-guided diffusion (DPD): per-category prototypes in latent space anchor hyperspherical semantic regions that guide the reverse denoising trajectory, and multiple candidates per prototype are ranked by a latent classifier's logit margin so only the most discriminative generated samples form the distilled set. Reports realistic, diverse and discriminative distilled samples across three high-resolution remote sensing scene-classification benchmarks.

2026-01

DGS — Difficulty-guided Sampling: Bridging the Target Gap between Dataset Distillation and Downstream Tasks

Mingzhuo Li, Guang Li, Linfeng Ye et al. · arXiv 2026paper ↗

Argues that existing distillation objectives match dataset-level statistics but ignore task-specific information needed for the downstream classifier, creating a target gap; introduces difficulty-guided sampling (DGS), a plug-in post-processing module that resamples an existing method's pool of distilled images to match a target difficulty distribution, and difficulty-aware guidance (DAG), which steers a diffusion generation process by the same difficulty signal. Reports consistent improvements when added to multiple existing distillation settings.

2026-01

SCG — Set-Coupled Guidance: Set-Level Coordination in Diffusion-Based Dataset Distillation

Ziang Gan, Qi Zhu, Libao Zhang · ICML 2026paper ↗code ↗

Proposes a plug-and-play auxiliary controller that coordinates diffusion sampling across an entire per-class image budget at once ('IPC-at-once') rather than image by image, injecting set-symmetric feedback at each denoising step so the generated set is coordinated as a group instead of independently sampled.

2026-01

Distribution shift in diffusion DD — Mitigating the Distribution Shift of Diffusion-based Dataset Distillation

Yue Xu, Chenyu Hu, Pengyu An et al. · CVPR 2026paper ↗

Identifies two distribution shifts in diffusion-based distillation: an optimal synthetic distribution should simplify, not replicate, the real data distribution given the synthetic set's small capacity, and the sampling process itself introduces a further deviation from the learned distribution; fixes both with an L1 sparsity regularizer during diffusion training and, at sampling time, a synchronous (non-sequential) denoising of the whole synthetic dataset with distribution regularizers.

2025-08

Enhancing Diffusion-based Dataset Distillation via Adversary-Guided Curriculum Sampling

Lexiao Zou, Gongwei Chen, Yanda Chen et al. · ICME 2025paper ↗

Argues that independently diffusion-sampled images lack diversity and create redundancy when aggregated into a distilled set, and proposes Adversary-guided Curriculum Sampling: partitions the distilled set into curricula, guiding each via an adversarial loss against a discriminator trained on already-sampled images to reduce overlap, with the discriminator's difficulty increasing across curricula for simple-to-complex coverage; reports a 4.1% gain on ImageWoof and 2.1% on ImageNet-1K over prior diffusion-based state of the art.

2025-08

Dataset Condensation with Color Compensation

Huyu Wu, Duo Su, Junjie Hou et al. · TMLR 2025paper ↗code ↗

After a calibrated real-image selection, fine-tunes a pretrained latent diffusion model on the condensed set to enhance the color diversity of the selected images rather than generating new ones from scratch, arguing color is both an information carrier and a semantic unit that prior condensation methods overlook; FID/IS results show the enhanced images avoid model collapse, and DC3 outperforms state-of-the-art methods across benchmarks.

2025-07

Label-Consistent Dataset Distillation with Detector-Guided Refinement

Yawen Zou, Guang Li, Zi Wang et al. · arXiv 2025paper ↗

Uses a detector trained on the original dataset to flag diffusion-generated synthetic images with label mismatches or low classification confidence, generates several diffusion candidates conditioned on the flagged image's prototype and label, and replaces it with the candidate that jointly maximizes detector confidence and dissimilarity to already-accepted synthetic samples. Reports state-of-the-art validation performance from the resulting gains in label consistency and image detail.

2025-07

Information-Guided Diffusion Sampling for Dataset Distillation

Linfeng Ye, Shayan Mohajer Hamidi, Guang Li et al. · NeurIPS 2025 Workshoppaper ↗

Steers diffusion-model sampling for dataset distillation by maximizing a combination of prototype information I(X;Y) and IPC-scaled contextual information H(X|Y), estimated with tractable variational lower bounds, to counter the low intra-class diversity of diffusion-sampled synthetic sets at small images-per-class; reports gains over existing diffusion-based distillation methods on Tiny-ImageNet and ImageNet subsets, especially at low IPC.

2025-07

Task-Specific Generative Dataset Distillation with Difficulty-Guided Sampling

Mingzhuo Li, Guang Li, Jiafeng Mao et al. · ICCV 2025 Workshoppaper ↗code ↗

Argues that prior generative distillation aligns the distilled set with the original dataset's overall distribution but ignores what the downstream classifier specifically needs, and proposes sampling from a larger generated image pool with a distribution matched to the original dataset's per-sample difficulty, applying a logarithmic transform to correct distributional bias, rather than matching representativeness alone.

2025-05

Diversity-Driven Generative Dataset Distillation Based on Diffusion Model with Self-Adaptive Memory

Mingzhuo Li, Guang Li, Jiafeng Mao et al. · ICIP 2025paper ↗

Uses a self-adaptive memory to measure how closely a diffusion-generated distilled set's distribution aligns with the real dataset, and feeds that alignment signal back into sampling to steer the diffusion model toward generating more diverse images during distillation. Reports outperforming existing generative dataset-distillation methods in most settings.

2025-05

CONCORD — CONCORD: Concept-Informed Diffusion for Dataset Distillation

Jianyang Gu, Haonan Wang, Ruoxi Jia et al. · arXiv 2025paper ↗code ↗

Retrieves fine-grained, distinguishable concepts from a large language model for each class label and feeds them into the diffusion denoising process to correct missing or incorrect object detail at the instance level, addressing a gap left by prior generative distillation methods that match only dataset-level statistics. Reports state-of-the-art results on ImageNet-1K and its subsets without relying on a pretrained classifier.

2024-08

Generative Dataset Distillation Based on Diffusion Model

Duo Su, Junjie Hou, Guang Li et al. · ECCV 2024 Workshoppaper ↗code ↗

Entry to the generative track of the ECCV 2024 Dataset Distillation Challenge: samples from SDXL-Turbo conditioned on class-name text prompts with post-generation augmentation, exploiting its speed to reach IPC=10 for Tiny-ImageNet and IPC=20 for CIFAR-100 within the challenge's 10-minute generation budget, versus IPC=1 for other diffusion entries; placed third in the track.

2024-06

InfoDist — Image Distillation for Safe Data Sharing in Histopathology

Zhe Li, Bernhard Kainz · MICCAI 2024paper ↗code ↗

Trains a latent diffusion model on histopathology data and builds a distilled, human-readable synthetic dataset by selecting the most informative generated images through graph community analysis of the representation space, addressing prior distillation methods' unreadable outputs and weak downstream performance. Reports downstream classifiers trained on the synthetic set reaching performance suitable for practical application, comparable to models trained on real data.