{"paper":{"title":"Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Tuna-2 shows that simple pixel patch embeddings can replace pretrained vision encoders for unified multimodal understanding and generation.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Belinda Zeng, Jonas Schult, Luke Zettlemoyer, Mengzhao Chen, Ping Luo, Sen He, Shoufa Chen, Tao Xiang, Tianhong Li, Weiming Ren, Wenhu Chen, Xiaoke Huang, Yatai Ji, Yuren Cong, Zhiheng Liu","submitted_at":"2026-04-27T17:59:56Z","abstract_excerpt":"Unified multimodal models typically rely on pretrained vision encoders and use separate visual representations for understanding and generation, creating misalignment between the two tasks and preventing fully end-to-end optimization from raw pixels. We introduce Tuna-2, a native unified multimodal model that performs visual understanding and generation directly based on pixel embeddings. Tuna-2 drastically simplifies the model architecture by employing simple patch embedding layers to encode visual input, completely discarding the modular vision encoder designs such as the VAE or the represen"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Tuna-2 achieves state-of-the-art performance in multimodal benchmarks, demonstrating that unified pixel-space modelling can fully compete with latent-space approaches for high-quality image generation. Moreover, while the encoder-based variant converges faster in early pretraining, Tuna-2's encoder-free design achieves stronger multimodal understanding at scale, particularly on tasks requiring fine-grained visual perception.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That simple patch embedding layers applied directly to pixels can extract sufficient visual features for both high-quality generation and fine-grained understanding without the inductive biases or pretraining provided by dedicated vision encoders.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Tuna-2 shows pixel embeddings can replace vision encoders in unified multimodal models, achieving competitive or superior results on understanding and generation benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Tuna-2 shows that simple pixel patch embeddings can replace pretrained vision encoders for unified multimodal understanding and generation.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"0d852a7d6154f4060b4655365e59ed1af7074e4b6ca131d8bfbdf770df1f3477"},"source":{"id":"2604.24763","kind":"arxiv","version":2},"verdict":{"id":"c5793468-ad57-400e-91e7-1d7890d12b31","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T04:27:47.732435Z","strongest_claim":"Tuna-2 achieves state-of-the-art performance in multimodal benchmarks, demonstrating that unified pixel-space modelling can fully compete with latent-space approaches for high-quality image generation. Moreover, while the encoder-based variant converges faster in early pretraining, Tuna-2's encoder-free design achieves stronger multimodal understanding at scale, particularly on tasks requiring fine-grained visual perception.","one_line_summary":"Tuna-2 shows pixel embeddings can replace vision encoders in unified multimodal models, achieving competitive or superior results on understanding and generation benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That simple patch embedding layers applied directly to pixels can extract sufficient visual features for both high-quality generation and fine-grained understanding without the inductive biases or pretraining provided by dedicated vision encoders.","pith_extraction_headline":"Tuna-2 shows that simple pixel patch embeddings can replace pretrained vision encoders for unified multimodal understanding and generation."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.24763/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-19T21:44:27.612842Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"fd1d77f395f8cfc7b6286c0bd81ecda604b883af1c59a7bd11bba05c71223451"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}