arXiv:2511.16213v1 Announce Type: new 
Abstract: Unsupervised image classification, or image clustering, aims to group unlabeled images into semantically meaningful categories. Early methods integrated representation learning and clustering within an iterative framework. However, the rise of foundational models have recently shifted focus solely to clustering, bypassing the representation learning step. In this work, we build upon a recent multi-head clustering approach by introducing adaptive nearest neighbor selection and cluster ensembling strategies to improve clustering performance. Our method, "Image Clustering through Cluster Ensembles" (ICCE), begins with a clustering stage, where we train multiple clustering heads on a frozen backbone, producing diverse image clusterings. We then employ a cluster ensembling technique to consolidate these potentially conflicting results into a unified consensus clustering. Finally, we train an image classifier using the consensus clustering result as pseudo-labels. ICCE achieves state-of-the-art performance on ten image classification benchmarks, achieving 99.3% accuracy on CIFAR10, 89% on CIFAR100, and 70.4% on ImageNet datasets, narrowing the performance gap with supervised methods. To the best of our knowledge, ICCE is the first fully unsupervised image classification method to exceed 70% accuracy on ImageNet.

تقدم الورقة نهجًا جديدًا لتصنيف الصور غير المراقب، يركز على تجميع الصور غير المعلّمة في فئات ذات معنى. تُحسّن الطريقة، المسماة Image Clustering through Cluster Ensembles (ICCE)، أداء التجميع من خلال دمج اختيار الجيران الأقرب التكيفي واستراتيجيات تجميع الكتل. يسمح هذا النهج بتدريب عدة رؤوس تجميع على هيكل ثابت، مما ينتج عنه تجميعات متنوعة يتم دمجها في تجميع إجماعي موحد.

El artículo presenta un enfoque novedoso para la clasificación de imágenes no supervisada, centrado en agrupar imágenes no etiquetadas en categorías significativas. El método, denominado Image Clustering through Cluster Ensembles (ICCE), mejora el rendimiento del agrupamiento al integrar la selección adaptativa de vecinos más cercanos y estrategias de ensamblaje de clústeres. Este enfoque permite entrenar múltiples cabezales de agrupamiento sobre un backbone fijo, produciendo agrupamientos diversos que se consolidan en un agrupamiento consensuado unificado.

Cet article présente une approche novatrice pour la classification d'images non supervisée, axée sur le regroupement d'images non étiquetées en catégories significatives. La méthode, nommée Image Clustering through Cluster Ensembles (ICCE), améliore les performances de regroupement en intégrant une sélection adaptative des plus proches voisins et des stratégies d'assemblage de clusters. Cette approche permet d'entraîner plusieurs têtes de regroupement sur un backbone fixe, produisant des regroupements divers qui sont consolidés en un regroupement consensuel unifié.

The paper presents a novel approach to unsupervised image classification, focusing on clustering unlabeled images into meaningful categories. The method, named Image Clustering through Cluster Ensembles (ICCE), enhances clustering performance by integrating adaptive nearest neighbor selection and cluster ensembling strategies. This approach allows for the training of multiple clustering heads on a fixed backbone, resulting in diverse clusterings that are consolidated into a unified consensus clustering.

Unsupervised Image Classification with Adaptive Nearest Neighbor Selection and Cluster Ensembles

arXiv:2401.17629v2 Announce Type: replace-cross 
Abstract: Diffusion models have recently emerged as a promising framework for Image Restoration (IR), owing to their ability to produce high-quality reconstructions and their compatibility with established methods. Existing methods for solving noisy inverse problems in IR, considers the pixel-wise data-fidelity. In this paper, we propose SaFaRI, a spatial-and-frequency-aware diffusion model for IR with Gaussian noise. Our model encourages images to preserve data-fidelity in both the spatial and frequency domains, resulting in enhanced reconstruction quality. We comprehensively evaluate the performance of our model on a variety of noisy inverse problems, including inpainting, denoising, and super-resolution. Our thorough evaluation demonstrates that SaFaRI achieves state-of-the-art performance on both the ImageNet datasets and FFHQ datasets, outperforming existing zero-shot IR methods in terms of LPIPS and FID metrics.

تقدم الورقة SaFaRI، وهو نموذج انتشار واعٍ للمساحة والتردد مصمم لاستعادة الصور (IR) والذي يتعامل بفعالية مع الضوضاء الغاوسية. يحسن هذا النموذج جودة إعادة البناء من خلال الحفاظ على دقة البيانات في كل من المجالات المكانية والترددية. تظهر التقييمات الشاملة أن SaFaRI يتفوق على الطرق الحالية لاستعادة الصور بدون إشراف على مجموعات بيانات ImageNet وFFHQ، محققًا أداءً رائدًا في مجموعة متنوعة من المشكلات العكسية المليئة بالضوضاء.

El artículo presenta a SaFaRI, un modelo de difusión consciente del espacio y la frecuencia diseñado para la restauración de imágenes (IR) que maneja eficazmente el ruido gaussiano. Este modelo mejora la calidad de reconstrucción al mantener la fidelidad de los datos en los dominios espacial y frecuencial. Evaluaciones exhaustivas demuestran que SaFaRI supera a los métodos IR de cero disparos existentes en los conjuntos de datos ImageNet y FFHQ, logrando un rendimiento de vanguardia en varios problemas inversos ruidosos.

Cet article présente SaFaRI, un modèle de diffusion conscient de l'espace et de la fréquence conçu pour la restauration d'images (IR) qui gère efficacement le bruit gaussien. Ce modèle améliore la qualité de reconstruction en maintenant la fidélité des données dans les domaines spatial et fréquentiel. Des évaluations complètes montrent que SaFaRI surpasse les méthodes IR zéro-shot existantes sur les ensembles de données ImageNet et FFHQ, atteignant des performances de pointe dans divers problèmes inverses bruyants.

The paper presents SaFaRI, a spatial-and-frequency-aware diffusion model designed for image restoration (IR) that effectively handles Gaussian noise. This model enhances reconstruction quality by maintaining data fidelity in both spatial and frequency domains. Comprehensive evaluations demonstrate that SaFaRI outperforms existing zero-shot IR methods on ImageNet and FFHQ datasets, achieving state-of-the-art performance in various noisy inverse problems.

Spatial-and-Frequency-aware Restoration method for Images based on Diffusion Models

arXiv:2511.16315v1 Announce Type: new 
Abstract: ImageNet-1K linear-probe transfer accuracy remains the default proxy for visual representation quality, yet it no longer predicts performance on scientific imagery. Across 46 modern vision model checkpoints, ImageNet top-1 accuracy explains only 34% of variance on ecology tasks and mis-ranks 30% of models above 75% accuracy. We present BioBench, an open ecology vision benchmark that captures what ImageNet misses. BioBench unifies 9 publicly released, application-driven tasks, 4 taxonomic kingdoms, and 6 acquisition modalities (drone RGB, web video, micrographs, in-situ and specimen photos, camera-trap frames), totaling 3.1M images. A single Python API downloads data, fits lightweight classifiers to frozen backbones, and reports class-balanced macro-F1 (plus domain metrics for FishNet and FungiCLEF); ViT-L models evaluate in 6 hours on an A6000 GPU. BioBench provides new signal for computer vision in ecology and a template recipe for building reliable AI-for-science benchmarks in any domain. Code and predictions are available at https://github.com/samuelstevens/biobench and results at https://samuelstevens.me/biobench.

تم تقديم BioBench كمعيار جديد للرؤية البيئية يتناول قيود ImageNet في التنبؤ بالأداء على الصور العلمية. يشمل 9 مهام مدفوعة بالتطبيقات، و4 ممالك تصنيفية، و6 طرق اكتساب، بإجمالي 3.1 مليون صورة. يهدف هذا المعيار إلى تحسين البحث البيئي من خلال توفير منصة موحدة لتقييم جودة التمثيل البصري في المهام البيئية.

BioBench se presenta como un nuevo estándar de visión ecológica que aborda las limitaciones de ImageNet en la predicción del rendimiento en imágenes científicas. Incluye 9 tareas impulsadas por aplicaciones, 4 reinos taxonómicos y 6 modalidades de adquisición, totalizando 3.1 millones de imágenes. Este estándar tiene como objetivo mejorar la investigación ecológica al proporcionar una plataforma unificada para evaluar la calidad de la representación visual en tareas ecológicas.

BioBench est présenté comme un nouvel étalon de vision écologique qui répond aux limites d'ImageNet dans la prédiction des performances sur les images scientifiques. Il englobe 9 tâches axées sur des applications, 4 règnes taxonomiques et 6 modalités d'acquisition, totalisant 3,1 millions d'images. Ce benchmark vise à améliorer la recherche écologique en fournissant une plateforme unifiée pour évaluer la qualité de la représentation visuelle dans les tâches écologiques.

BioBench is introduced as an open ecology vision benchmark that addresses the limitations of ImageNet in predicting performance on scientific imagery. It encompasses 9 application-driven tasks, 4 taxonomic kingdoms, and 6 acquisition modalities, totaling 3.1 million images. The benchmark aims to enhance ecological research by providing a unified platform for evaluating visual representation quality in ecological tasks.

BioBench: A Blueprint to Move Beyond ImageNet for Scientific ML Benchmarks

arXiv:2511.16546v1 Announce Type: new 
Abstract: Visual Auto-Regressive (VAR) models significantly reduce inference steps through the "next-scale" prediction paradigm. However, progressive multi-scale generation incurs substantial memory overhead due to cumulative KV caching, limiting practical deployment.
  We observe a scale-depth asymmetric dependency in VAR: early scales exhibit extreme sensitivity to network depth, while later scales remain robust to depth reduction. Inspired by this, we propose VARiant: by equidistant sampling, we select multiple subnets ranging from 16 to 2 layers from the original 30-layer VAR-d30 network. Early scales are processed by the full network, while later scales utilize subnet. Subnet and the full network share weights, enabling flexible depth adjustment within a single model.
  However, weight sharing between subnet and the entire network can lead to optimization conflicts. To address this, we propose a progressive training strategy that breaks through the Pareto frontier of generation quality for both subnets and the full network under fixed-ratio training, achieving joint optimality.
  Experiments on ImageNet demonstrate that, compared to the pretrained VAR-d30 (FID 1.95), VARiant-d16 and VARiant-d8 achieve nearly equivalent quality (FID 2.05/2.12) while reducing memory consumption by 40-65%. VARiant-d2 achieves 3.5 times speedup and 80% memory reduction at moderate quality cost (FID 2.97). In terms of deployment, VARiant's single-model architecture supports zero-cost runtime depth switching and provides flexible deployment options from high quality to extreme efficiency, catering to diverse application scenarios.

تقدم الورقة نهجًا مبتكرًا لنمذجة الانحدار الذاتي البصري (VAR) من خلال تقديم VARiant، الذي يحسن استخدام الذاكرة عبر استراتيجيات التدريب التدريجي. تسمح هذه الطريقة بتعديلات مرنة في العمق داخل الشبكة، مما يعالج قيود التوليد متعدد المقاييس التقليدي. من خلال معالجة المقاييس المبكرة باستخدام شبكة كاملة والمقاييس اللاحقة باستخدام الشبكات الفرعية، يعزز VARiant الكفاءة مع الحفاظ على الأداء.

El artículo presenta un enfoque novedoso para la modelización autorregresiva visual (VAR), introduciendo VARiant, que optimiza el uso de memoria mediante estrategias de entrenamiento progresivo. Este método permite ajustes flexibles de profundidad en la red, abordando las limitaciones de la generación multiescalar tradicional. Al procesar las escalas tempranas con una red completa y las escalas posteriores con subredes, VARiant mejora la eficiencia mientras mantiene el rendimiento.

Cet article présente une approche novatrice pour la modélisation auto-régressive visuelle (VAR), introduisant VARiant, qui optimise l'utilisation de la mémoire grâce à des stratégies d'entraînement progressif. Cette méthode permet des ajustements de profondeur flexibles dans le réseau, répondant aux limitations de la génération multi-échelle traditionnelle. En traitant les premières échelles avec un réseau complet et les échelles ultérieures avec des sous-réseaux, VARiant améliore l'efficacité tout en maintenant la performance.

The paper presents a novel approach to Visual Auto-Regressive (VAR) modeling, introducing VARiant, which optimizes memory usage by employing progressive training strategies. This method allows for flexible depth adjustments in the network, addressing the limitations of traditional multi-scale generation. By processing early scales with a full network and later scales with subnets, VARiant enhances efficiency while maintaining performance.

Unsupervised Image Classification with Adaptive Nearest Neighbor Selection and Cluster Ensembles

Was this article worth reading? Share it