{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:QIGCND7F5EA5UUZFFISGBCF4LU","short_pith_number":"pith:QIGCND7F","schema_version":"1.0","canonical_sha256":"820c268fe5e901da53252a246088bc5d17b86e8743f2d57332aefa6ff59f0df3","source":{"kind":"arxiv","id":"2605.13974","version":1},"attestation_state":"computed","paper":{"title":"Few Channels Draw The Whole Picture: Revealing Massive Activations in Diffusion Transformers","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"A small set of massive activation channels in Diffusion Transformers controls image semantics in function, space, and transfer.","cross_cats":["cs.AI","cs.MM"],"primary_cat":"cs.CV","authors_text":"Davide Bucciarelli, Evelyn Turri, Lorenzo Baraldi, Marcella Cornia, Sara Sarto","submitted_at":"2026-05-13T18:00:12Z","abstract_excerpt":"Diffusion Transformers (DiTs) and related flow-based architectures are now among the strongest text-to-image generators, yet the internal mechanisms through which prompts shape image semantics remain poorly understood. In this work, we study massive activations: a small subset of hidden-state channels whose responses are consistently much larger than the rest. We show that, despite their sparsity, these few channels effectively draw the whole picture, in three complementary senses. First, they are functionally critical: a controlled disruption probe that zeroes the massive channels causes a sh"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.13974","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-13T18:00:12Z","cross_cats_sorted":["cs.AI","cs.MM"],"title_canon_sha256":"ebcb4ed9a826ce4fa6a782d6e769fc40ec2ea6c98664c13101099536b6fae800","abstract_canon_sha256":"72e124ef19189a9715e1e702a5809224fc9ed0ca6baee97b9179e5ebe6338d0e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:13.459190Z","signature_b64":"LBWM0SW51LKwiPVr4drIyApFg0L3kfMve2HZ6dd8KZoXfL8BuIpE25SwDYz9o5o2H6ozMcFbxz3Ttt12JvAAAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"820c268fe5e901da53252a246088bc5d17b86e8743f2d57332aefa6ff59f0df3","last_reissued_at":"2026-05-17T23:39:13.458382Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:13.458382Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Few Channels Draw The Whole Picture: Revealing Massive Activations in Diffusion Transformers","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"A small set of massive activation channels in Diffusion Transformers controls image semantics in function, space, and transfer.","cross_cats":["cs.AI","cs.MM"],"primary_cat":"cs.CV","authors_text":"Davide Bucciarelli, Evelyn Turri, Lorenzo Baraldi, Marcella Cornia, Sara Sarto","submitted_at":"2026-05-13T18:00:12Z","abstract_excerpt":"Diffusion Transformers (DiTs) and related flow-based architectures are now among the strongest text-to-image generators, yet the internal mechanisms through which prompts shape image semantics remain poorly understood. In this work, we study massive activations: a small subset of hidden-state channels whose responses are consistently much larger than the rest. We show that, despite their sparsity, these few channels effectively draw the whole picture, in three complementary senses. First, they are functionally critical: a controlled disruption probe that zeroes the massive channels causes a sh"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"despite their sparsity, these few channels effectively draw the whole picture, in three complementary senses: functionally critical, spatially organized, and transferable.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the identification of massive channels via magnitude statistics is stable across prompts and models and that the controlled disruption probe isolates their causal role without confounding effects from the rest of the network dynamics.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A sparse set of massive activation channels in DiTs carries semantic information, proven critical by disruption tests, spatially aligned with image subjects via clustering, and transferable for semantic interpolation between prompts.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A small set of massive activation channels in Diffusion Transformers controls image semantics in function, space, and transfer.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"73d37097b53bb84eb28427dbdd8bc4a5be8d331d1bcdde6e963c49e4481e374c"},"source":{"id":"2605.13974","kind":"arxiv","version":1},"verdict":{"id":"7369825b-e2ae-4aa8-872a-9744e0a47945","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T05:55:17.889017Z","strongest_claim":"despite their sparsity, these few channels effectively draw the whole picture, in three complementary senses: functionally critical, spatially organized, and transferable.","one_line_summary":"A sparse set of massive activation channels in DiTs carries semantic information, proven critical by disruption tests, spatially aligned with image subjects via clustering, and transferable for semantic interpolation between prompts.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the identification of massive channels via magnitude statistics is stable across prompts and models and that the controlled disruption probe isolates their causal role without confounding effects from the rest of the network dynamics.","pith_extraction_headline":"A small set of massive activation channels in Diffusion Transformers controls image semantics in function, space, and transfer."},"references":{"count":41,"sample":[{"doi":"","year":2023,"title":"Building normalizing flows with stochastic interpolants","work_id":"1351da88-cc93-4497-97a3-5eb936b2682c","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"All are Worth Words: A ViT Backbone for Diffusion Models","work_id":"6209d5f2-cc4d-455e-9d7e-d08e1bbaff0c","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Tiny Inference-Time Scaling with Latent Verifiers","work_id":"da382efe-e434-4a39-b51b-9c08cd9d88e5","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation","work_id":"0a9f6fd1-d3f5-4778-b034-3ef328f69247","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Vision Transformers Need Registers","work_id":"9ad78f6d-8598-4484-9bd7-752fdb907474","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":41,"snapshot_sha256":"a26edf58e0f317b5e80b62d0423e2b2a1b2f7e86589869c29f01046f92cffe00","internal_anchors":4},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2605.13974","created_at":"2026-05-17T23:39:13.458465+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.13974v1","created_at":"2026-05-17T23:39:13.458465+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13974","created_at":"2026-05-17T23:39:13.458465+00:00"},{"alias_kind":"pith_short_12","alias_value":"QIGCND7F5EA5","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"QIGCND7F5EA5UUZF","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"QIGCND7F","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/QIGCND7F5EA5UUZFFISGBCF4LU","json":"https://pith.science/pith/QIGCND7F5EA5UUZFFISGBCF4LU.json","graph_json":"https://pith.science/api/pith-number/QIGCND7F5EA5UUZFFISGBCF4LU/graph.json","events_json":"https://pith.science/api/pith-number/QIGCND7F5EA5UUZFFISGBCF4LU/events.json","paper":"https://pith.science/paper/QIGCND7F"},"agent_actions":{"view_html":"https://pith.science/pith/QIGCND7F5EA5UUZFFISGBCF4LU","download_json":"https://pith.science/pith/QIGCND7F5EA5UUZFFISGBCF4LU.json","view_paper":"https://pith.science/paper/QIGCND7F","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.13974&json=true","fetch_graph":"https://pith.science/api/pith-number/QIGCND7F5EA5UUZFFISGBCF4LU/graph.json","fetch_events":"https://pith.science/api/pith-number/QIGCND7F5EA5UUZFFISGBCF4LU/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/QIGCND7F5EA5UUZFFISGBCF4LU/action/timestamp_anchor","attest_storage":"https://pith.science/pith/QIGCND7F5EA5UUZFFISGBCF4LU/action/storage_attestation","attest_author":"https://pith.science/pith/QIGCND7F5EA5UUZFFISGBCF4LU/action/author_attestation","sign_citation":"https://pith.science/pith/QIGCND7F5EA5UUZFFISGBCF4LU/action/citation_signature","submit_replication":"https://pith.science/pith/QIGCND7F5EA5UUZFFISGBCF4LU/action/replication_record"}},"created_at":"2026-05-17T23:39:13.458465+00:00","updated_at":"2026-05-17T23:39:13.458465+00:00"}