{"paper":{"title":"CoCa: Contrastive Captioners are Image-Text Foundation Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"CoCa jointly trains contrastive and captioning losses in one encoder-decoder to create image-text foundation models that reach new state-of-the-art on ImageNet and multimodal tasks.","cross_cats":["cs.LG","cs.MM"],"primary_cat":"cs.CV","authors_text":"Jiahui Yu, Legg Yeung, Mojtaba Seyedhosseini, Vijay Vasudevan, Yonghui Wu, Zirui Wang","submitted_at":"2022-05-04T07:01:14Z","abstract_excerpt":"Exploring large-scale pretrained foundation models is of significant interest in computer vision because these models can be quickly transferred to many downstream tasks. This paper presents Contrastive Captioner (CoCa), a minimalist design to pretrain an image-text encoder-decoder foundation model jointly with contrastive loss and captioning loss, thereby subsuming model capabilities from contrastive approaches like CLIP and generative methods like SimVLM. In contrast to standard encoder-decoder transformers where all decoder layers attend to encoder outputs, CoCa omits cross-attention in the"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"CoCa obtains 86.3% zero-shot top-1 accuracy on ImageNet, 90.6% with a frozen encoder and learned classification head, and new state-of-the-art 91.0% top-1 accuracy on ImageNet with a finetuned encoder, while also leading on Kinetics, MSCOCO, VQA, and other tasks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That omitting cross-attention in the first half of the decoder layers cleanly separates unimodal text representations from multimodal ones without harming overall capacity or optimization stability.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"CoCa unifies contrastive and generative pretraining in one image-text model to reach 86.3% zero-shot ImageNet accuracy and new state-of-the-art results on multiple downstream benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"CoCa jointly trains contrastive and captioning losses in one encoder-decoder to create image-text foundation models that reach new state-of-the-art on ImageNet and multimodal tasks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"62f230de0b1498c62f4242da007033c4a6474eba7ea984251c75c03889628f78"},"source":{"id":"2205.01917","kind":"arxiv","version":2},"verdict":{"id":"1f439bdc-a6c5-418f-9c37-c73563009d9e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T10:48:13.746831Z","strongest_claim":"CoCa obtains 86.3% zero-shot top-1 accuracy on ImageNet, 90.6% with a frozen encoder and learned classification head, and new state-of-the-art 91.0% top-1 accuracy on ImageNet with a finetuned encoder, while also leading on Kinetics, MSCOCO, VQA, and other tasks.","one_line_summary":"CoCa unifies contrastive and generative pretraining in one image-text model to reach 86.3% zero-shot ImageNet accuracy and new state-of-the-art results on multiple downstream benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That omitting cross-attention in the first half of the decoder layers cleanly separates unimodal text representations from multimodal ones without harming overall capacity or optimization stability.","pith_extraction_headline":"CoCa jointly trains contrastive and captioning losses in one encoder-decoder to create image-text foundation models that reach new state-of-the-art on ImageNet and multimodal tasks."},"references":{"count":80,"sample":[{"doi":"","year":2021,"title":"On the Opportunities and Risks of Foundation Models","work_id":"a18039e9-928d-47c9-a836-32656a71bf71","ref_index":1,"cited_arxiv_id":"2108.07258","is_internal_anchor":true},{"doi":"","year":2018,"title":"BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding","work_id":"ed240a10-5b19-406c-baa5-30803f465785","ref_index":2,"cited_arxiv_id":"1810.04805","is_internal_anchor":true},{"doi":"","year":1910,"title":"Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer","work_id":"50e3b368-0243-4726-8186-233869802ad1","ref_index":3,"cited_arxiv_id":"1910.10683","is_internal_anchor":true},{"doi":"","year":1901,"title":"Language models are few-shot learners","work_id":"04bc68bc-b7df-4ec1-8599-da037bd4f085","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Thekkath, and Yonghui Wu","work_id":"6192fe03-df78-42e3-88fa-002af7197669","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":80,"snapshot_sha256":"15f6ce8074fb38f0cdc4f7262d52056ff86535f1620fd10356acbefe97446259","internal_anchors":15},"formal_canon":{"evidence_count":2,"snapshot_sha256":"ef2818b7fe9617f272021a3d4e2a136215e8d7bade34bf7398fd12bf60a5724f"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}