{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/cosmos-cross-modality-self-distillation-for","title":"COSMOS: Cross-Modality Self-Distillation for Vision Language Pre-training","arxiv_id":"2412.01814","date":"2024-12-02","proceeding":"CVPR 2025 1","authors":["Sanghwan Kim","Rui Xiao","Mariana-Iuliana Georgescu","Stephan Alaniz","Zeynep Akata"],"abstract":"Vision-Language Models (VLMs) trained with contrastive loss have achieved significant advancements in various vision and language tasks. However, the global nature of the contrastive loss makes VLMs focus predominantly on foreground objects, neglecting other crucial information in the image, which limits their effectiveness in downstream tasks. To address these challenges, we propose COSMOS: CrOSs-MOdality Self-distillation for vision-language pre-training that integrates a novel text-cropping strategy and cross-attention module into a self-supervised learning framework. We create global and local views of images and texts (i.e., multi-modal augmentations), which are essential for self-distillation in VLMs. We further introduce a cross-attention module, enabling COSMOS to learn comprehensive cross-modal representations optimized via a cross-modality self-distillation loss. COSMOS consistently outperforms previous strong baselines on various zero-shot downstream tasks, including retrieval, classification, and semantic segmentation. Additionally, it surpasses CLIP-based models trained on larger datasets in visual perception and contextual understanding tasks. Code is available at https://github.com/ExplainableML/cosmos.","url_abs":"https://arxiv.org/abs/2412.01814v2","url_pdf":"https://arxiv.org/pdf/2412.01814v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"cosmos-cross-modality-self-distillation-for","repo_url":"https://github.com/ExplainableML/cosmos","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}}],"tasks":[{"task_slug":"self-supervised-learning","task_name":"Self-Supervised Learning"},{"task_slug":"semantic-segmentation","task_name":"Semantic Segmentation"},{"task_slug":"unsupervised-semantic-segmentation-with","task_name":"Unsupervised Semantic Segmentation with Language-image Pre-training"},{"task_slug":"zero-shot-segmentation","task_name":"Zero Shot Segmentation"},{"task_slug":"zero-shot-cross-modal-retrieval","task_name":"Zero-Shot Cross-Modal Retrieval"}],"methods":[{"method_slug":"concatenated-skip-connection","method_name":"Concatenated Skip Connection"},{"method_slug":"focus","method_name":"Focus"},{"method_slug":"softmax","method_name":"Softmax"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/unsupervised-semantic-segmentation-with-4","task":"Unsupervised Semantic Segmentation with Language-image Pre-training","dataset":"ADE20K","model":"COSMOS ViT-B/16","rank_in_archive_order":5,"of":13,"metrics":{"Mean IoU (val)":"17.7"},"uses_additional_data":false},{"leaderboard":"/sota/unsupervised-semantic-segmentation-with-10","task":"Unsupervised Semantic Segmentation with Language-image Pre-training","dataset":"COCO-Object","model":"COSMOS ViT-B/16","rank_in_archive_order":8,"of":12,"metrics":{"mIoU":"31.3"},"uses_additional_data":false},{"leaderboard":"/sota/unsupervised-semantic-segmentation-with-9","task":"Unsupervised Semantic Segmentation with Language-image Pre-training","dataset":"COCO-Stuff-171","model":"COSMOS ViT-B/16","rank_in_archive_order":7,"of":12,"metrics":{"mIoU":"23.2"},"uses_additional_data":false},{"leaderboard":"/sota/unsupervised-semantic-segmentation-with-3","task":"Unsupervised Semantic Segmentation with Language-image Pre-training","dataset":"Cityscapes val","model":"COSMOS ViT-B/16","rank_in_archive_order":4,"of":12,"metrics":{"mIoU":"34.7"},"uses_additional_data":false},{"leaderboard":"/sota/unsupervised-semantic-segmentation-with-8","task":"Unsupervised Semantic Segmentation with Language-image Pre-training","dataset":"PASCAL Context-59","model":"COSMOS ViT-B/16","rank_in_archive_order":8,"of":12,"metrics":{"mIoU":"33.7"},"uses_additional_data":false},{"leaderboard":"/sota/unsupervised-semantic-segmentation-with-7","task":"Unsupervised Semantic Segmentation with Language-image Pre-training","dataset":"PascalVOC-20","model":"COSMOS ViT-B/16","rank_in_archive_order":8,"of":10,"metrics":{"mIoU":"77.7"},"uses_additional_data":false},{"leaderboard":"/sota/zero-shot-segmentation-on-ade20k-training","task":"Zero Shot Segmentation","dataset":"ADE20K training-free zero-shot segmentation","model":"COSMOS ViT-B/16","rank_in_archive_order":1,"of":5,"metrics":{"mIoU":"17.7"},"uses_additional_data":false},{"leaderboard":"/sota/zero-shot-cross-modal-retrieval-on-coco-2014","task":"Zero-Shot Cross-Modal Retrieval","dataset":"COCO 2014","model":"COSMOS ViT-B/16","rank_in_archive_order":8,"of":18,"metrics":{"Image-to-text R@1":"68.0","Image-to-text R@10":"92.5","Image-to-text R@5":"87.8","Text-to-image R@1":"52.5","Text-to-image R@10":"84.9","Text-to-image R@5":"77.2"},"uses_additional_data":false},{"leaderboard":"/sota/zero-shot-cross-modal-retrieval-on-coco-2014","task":"Zero-Shot Cross-Modal Retrieval","dataset":"COCO 2014","model":"COSMOS ViT-B/32","rank_in_archive_order":12,"of":18,"metrics":{"Image-to-text R@1":"64.3","Image-to-text R@10":"92.0","Image-to-text R@5":"86.5","Text-to-image R@1":"48.4","Text-to-image R@10":"82.6","Text-to-image R@5":"74.2"},"uses_additional_data":false},{"leaderboard":"/sota/zero-shot-cross-modal-retrieval-on-flickr30k","task":"Zero-Shot Cross-Modal Retrieval","dataset":"Flickr30k","model":"COSMOS ViT-B/16","rank_in_archive_order":4,"of":22,"metrics":{"Image-to-text R@1":"92.9","Image-to-text R@10":"99.9","Image-to-text R@5":"99.4","Text-to-image R@1":"80.3","Text-to-image R@10":"97.6","Text-to-image R@5":"95.3"},"uses_additional_data":false},{"leaderboard":"/sota/zero-shot-cross-modal-retrieval-on-flickr30k","task":"Zero-Shot Cross-Modal Retrieval","dataset":"Flickr30k","model":"COSMOS ViT-B/32","rank_in_archive_order":11,"of":22,"metrics":{"Image-to-text R@1":"89.9","Image-to-text R@10":"99.3","Image-to-text R@5":"98.8","Text-to-image R@1":"76.1","Text-to-image R@10":"96.2","Text-to-image R@5":"92.8"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=2412.01814","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}