Self-Supervised Representation-Guided Generative Dataset Distillation

CoRe-Diffusion: Bridging the Generative-Discriminative Gap in Dataset Distillation via Manifold-Aligned Contrastive Guidance

Dataset Distillation aims to condense large-scale datasets into a tiny but highly informative subset, enabling efficient training while achieving performance comparable to the full dataset. To overcome the scalability bottlenecks of traditional optimization, Generative Dataset Distillation (GDD) leverages diffusion priors to reformulate the dataset condensation process as an efficient generative sampling paradigm. However, existing GDD paradigms primarily optimize intra-class likelihood while largely overlooking explicit inter-class separation, which may cause generated features to concentrate near decision boundaries. Furthermore, relying on manually designed guidance targets to steer synthesis often pushes the diffusion trajectory off the natural data manifold, inducing severe structural artifacts. To address these issues, CoRe-Diffusion is proposed as a unified framework. Specifically, it introduces Contrastive Negative Guidance, which utilizes inter-class mode centers—computed but largely unused during synthesis by existing methods—to exert zero-overhead repulsive gradients against hard negatives. To preserve synthesis fidelity, Manifold-Aligned Real-Anchor Discovery strictly constrains the guidance targets to the exact latent representations of real images, preventing the trajectory from drifting. This intrinsically preserves complex semantic structures, bypassing the prohibitive computational overhead of relying on external generative priors or auxiliary modules. Alongside an Annealed Sampling Schedule for smooth trajectory evolution, CoRe-Diffusion achieves state-of-the-art performance, yielding a 3.8\% absolute accuracy improvement on ImageNet-1K while introducing negligible computational overhead.

DeCO: Discriminative Evidence Composition for Fine-Grained Dataset Distillation

Dataset distillation compresses a large training set into a compact synthetic set while preserving its downstream utility. However, existing methods primarily preserve global image statistics and may overlook the localized evidence essential for fine-grained visual classification (FGVC), such as object parts, subtle textures, and region-specific structures. We formulate fine-grained dataset distillation as budgeted discriminative-evidence preservation and propose Discriminative Evidence Composition (DeCO). DeCO uses attention rollout from a pretrained TransFG teacher to identify informative patches, applies spatial diversification to reduce redundant coverage, and organizes the resulting regions into class-wise evidence banks. Multiple same-class regions are then packed into compact grid-composed images. The teacher is used only for dataset construction, whereas downstream students are trained with standard hard-label supervision without teacher logits. Experiments on CUB-200-2011, FGVC-Aircraft, and Stanford Cars show that DeCO consistently outperforms representative coreset and dataset-distillation baselines under different IPC budgets.

Dataset Distillation Based on Saliency-Driven Prototype Alignment

Dataset distillation aims to synthesize compact datasets that can approximate the performance of full-data training while significantly reducing computational and storage costs. However, diffusion-based distillation methods often struggle to preserve structural coherence and generalization, especially in visually complex domains. This issue often stems from latent prototypes that are weakly aligned with class-discriminative regions and contaminated by irrelevant background, thereby degrading generation quality and generalization. To address this limitation, we propose a saliency-driven distillation framework that constructs class-discriminative latent prototypes to enhance representativeness and generalization. The framework proceeds in two stages: (1) ensemble Grad-CAM++ saliency is used to construct prototypes emphasizing class-discriminative regions, and (2) hard-prototype refinement is then applied to construct challenging yet class-consistent prototypes, thereby enhancing discriminability and diversity. Importantly, the diffusion backbones (e.g., LDM and DiT) remain frozen; only lightweight classifiers used for saliency extraction are trained. Extensive experiments across multiple benchmarks demonstrate consistent performance improvements over strong baselines. Code will be released.

Enhancing self-supervised representation learning through lightweight learnable data augmentation

Data augmentation plays a central role in self-supervised learning, as the quality and diversity of augmented views strongly influence the learned representations. However, most existing self-supervised methods rely on fixed stochastic augmentation pipelines, while more adaptive alternatives often require expensive policy search, adversarial training, or additional optimization procedures. In this paper, we propose a lightweight learnable augmentation framework based on Extreme Learning Machines (ELM) for self-supervised visual representation learning. The proposed module predicts image-dependent transformation parameters and applies them through a differentiable augmentation operator, enabling joint optimization with the representation model while introducing minimal additional computational overhead. The framework is integrated into three representative self-supervised learning methods: SimCLR, BYOL, and SimSiam. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny ImageNet show that the proposed method consistently improves linear evaluation performance relative to reproduced baselines across most settings. In particular, the method yields notable gains on CIFAR datasets and remains effective on the more challenging Tiny ImageNet benchmark. A per-class difficulty analysis further shows that the proposed augmentation strategy substantially improves performance on hard classes, indicating stronger robustness to challenging categories while maintaining competitive overall performance. In general, the results demonstrate that lightweight learnable augmentation can effectively enhance self-supervised representation learning across different frameworks and datasets.

Isotropic Embedding Perturbations for Robust Vision Language Encoders

Data augmentation is fundamental to training modern deep vision and multimodal models. While individual methods, such as RandAug, CutMix, Mixup, RandErase, and DropPath, offer strong regularization effects, their combined use has saturated in performance due to overlapping functionalities, and aggressive pixel-level manipulations may disrupt delicate cross-modal alignment. This saturation motivates the search for a new augmentation axis within the embedding space rather than the input space. We introduce Aether, a simple plug-in method that applies diffusion-style random perturbations in the embedding space via controlled alpha-mixing, specifically designed to provide isotropic regularization that remains semantically consistent. Inspired by feature-space perturbations in language models and image degradation in generative pretraining, Aether induces mild yet effective perturbations that smooth the representations without compromising the fine-grained structural information required for strong vision-language encoders. Across diverse architectures and across multiple recognition tasks, Aether delivers consistent gains over the advanced recipe combining CutMix, Mixup, DropPath, and RandAug---a level of improvement rarely observed with modern augmentation alternatives. Notably, Aether demonstrates superior effectiveness in multi-modal alignment, succeeding where traditional pixel-space augmentations fail by providing a stable, isotropic regularization signal that respects the integrity of the high-dimensional feature space.

Self-distillation framework via historical collaborative learning for image classification