{"paper":{"title":"Zero-Shot Text-to-Image Generation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A transformer that models text and image tokens as one autoregressive stream achieves competitive zero-shot text-to-image generation at sufficient scale.","cross_cats":["cs.LG"],"primary_cat":"cs.CV","authors_text":"Aditya Ramesh, Alec Radford, Chelsea Voss, Gabriel Goh, Ilya Sutskever, Mark Chen, Mikhail Pavlov, Scott Gray","submitted_at":"2021-02-24T06:42:31Z","abstract_excerpt":"Text-to-image generation has traditionally focused on finding better modeling assumptions for training on a fixed dataset. These assumptions might involve complex architectures, auxiliary losses, or side information such as object part labels or segmentation masks supplied during training. We describe a simple approach for this task based on a transformer that autoregressively models the text and image tokens as a single stream of data. With sufficient data and scale, our approach is competitive with previous domain-specific models when evaluated in a zero-shot fashion."},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"With sufficient data and scale, our approach is competitive with previous domain-specific models when evaluated in a zero-shot fashion.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That simply increasing model size and data volume will continue to close the gap to specialized models without introducing new failure modes or requiring additional inductive biases.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A transformer autoregressively models text and image tokens as one stream and produces competitive zero-shot text-to-image results at sufficient scale.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A transformer that models text and image tokens as one autoregressive stream achieves competitive zero-shot text-to-image generation at sufficient scale.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"13db172eac172da28dfbbecbfd58e7c97ef70b1dabb61fe7a9acc90130fd22bc"},"source":{"id":"2102.12092","kind":"arxiv","version":2},"verdict":{"id":"d06740f6-2e1e-4b26-9524-ef4eedfd647c","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T22:22:00.876735Z","strongest_claim":"With sufficient data and scale, our approach is competitive with previous domain-specific models when evaluated in a zero-shot fashion.","one_line_summary":"A transformer autoregressively models text and image tokens as one stream and produces competitive zero-shot text-to-image results at sufficient scale.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That simply increasing model size and data volume will continue to close the gap to specialized models without introducing new failure modes or requiring additional inductive biases.","pith_extraction_headline":"A transformer that models text and image tokens as one autoregressive stream achieves competitive zero-shot text-to-image generation at sufficient scale."},"references":{"count":23,"sample":[{"doi":"","year":2015,"title":"Bowman et al","work_id":"52301570-311d-400c-a35e-e91c2dafc161","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Using a linear annealing schedule for this typically led to divergence","work_id":"a760669a-0fff-48d1-8242-87dac551d654","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"row, column, row, row","work_id":"0336e6b0-451d-47dc-8d97-1a46d4107b72","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Our model uses 128 gradient scales, one for each of its resblocks","work_id":"acbee4e8-41fd-48e1-b387-4b6685b9e6e2","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"In particular, store all gains, biases, embeddings, and unembeddings in 32-bit precision, with 32-bit gradients (including for remote communication) and 32-bit Adam moments","work_id":"2cbd3bc9-f9b0-4459-902d-4ad616b9d452","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":23,"snapshot_sha256":"20e3e9821b8d8cdb6d7d5423026327248332c552ac100fa4abd9e4fd6b4d7666","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"d12df4dfee20ea1cf31da99d9d8fe14ff7462fb2fce8615d7054987eb1475f96"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}