{"paper":{"title":"HeatKV: Head-tuned KV-cache Compression for Visual Autoregressive Modeling","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Head-specific attention ranking doubles KV-cache compression in visual autoregressive image models while preserving quality.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Axel Berg, Chuteng Zhou, Durmus Alp Emre Acar, Jonathan Cederlund, Pontus Giselsson","submitted_at":"2026-05-14T14:22:34Z","abstract_excerpt":"Visual Autoregressive (VAR) models have recently demonstrated impressive image generation quality while maintaining low latency. However, they suffer from severe KV-cache memory constraints, often requiring gigabytes of memory per generated image. We introduce HeatKV, a novel compression method that adapts cache allocation in each head based on its attention to previously generated scales. Using a small offline calibration set, the attention heads are ranked according to their attention scores over prior scales. Based on this ranking, we construct a static pruning schedule tailored to a given "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Applied to the Infinity-2B model, HeatKV achieves 2× higher compression ratio in memory allocation for KV cache compared to existing methods, while maintaining similar or better image fidelity, prompt alignment and human perception score.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That a static pruning schedule derived from attention scores on a small offline calibration set will generalize to arbitrary prompts and generation lengths without measurable quality loss.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"HeatKV ranks attention heads by their focus on prior scales using offline calibration data and applies a static per-head pruning schedule, delivering 2x higher KV-cache compression than prior methods on the Infinity-2B model with comparable image fidelity.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Head-specific attention ranking doubles KV-cache compression in visual autoregressive image models while preserving quality.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"65ba968a434d962722aa0603e36d76022c7c8a7be8960fab444d48979d528093"},"source":{"id":"2605.14877","kind":"arxiv","version":1},"verdict":{"id":"fb93b85d-6058-4875-89dc-c22b53b3a353","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T03:15:13.332676Z","strongest_claim":"Applied to the Infinity-2B model, HeatKV achieves 2× higher compression ratio in memory allocation for KV cache compared to existing methods, while maintaining similar or better image fidelity, prompt alignment and human perception score.","one_line_summary":"HeatKV ranks attention heads by their focus on prior scales using offline calibration data and applies a static per-head pruning schedule, delivering 2x higher KV-cache compression than prior methods on the Infinity-2B model with comparable image fidelity.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That a static pruning schedule derived from attention scores on a small offline calibration set will generalize to arbitrary prompts and generation lengths without measurable quality loss.","pith_extraction_headline":"Head-specific attention ranking doubles KV-cache compression in visual autoregressive image models while preserving quality."},"references":{"count":40,"sample":[{"doi":"","year":2025,"title":"Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis,","work_id":"977492f9-1add-4df5-aa71-c2e7f1a694b6","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2014,"title":"I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,”Advances in neural information processing systems, vol. 27, ","work_id":"4b8320ae-4a24-4686-85ca-93e608f3af08","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"A style-based generator architecture for generative adversarial networks,","work_id":"db3bde24-03b8-4609-98c5-e9672550b362","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Stylegan-xl: Scaling stylegan to large diverse datasets,","work_id":"f5a2bd2d-af70-447d-a377-f8717494b10f","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Diffusion models beat gans on image synthesis,","work_id":"35493ec7-8582-4e9b-a3dc-d835aca70a28","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":40,"snapshot_sha256":"bb5685a19cb18a9b4eee104188b34d8e9606abdd26d8b09aefd67fe8dfe4d354","internal_anchors":2},"formal_canon":{"evidence_count":2,"snapshot_sha256":"63aeb8692867d4ce5341a164139d63f584be7c504aea060e9cd9ca6f67647c72"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}