{"paper":{"title":"Kiwi-Edit: Versatile Video Editing via Instruction and Reference Guidance","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Kiwi-Edit achieves state-of-the-art results in controllable video editing by combining instructions with reference images through a new data pipeline and architecture.","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Guoqiang Liang, Mike Zheng Shou, Yanzhe Chen, Yiqi Lin, Zechen Bai, Ziyun Zeng","submitted_at":"2026-03-02T18:46:28Z","abstract_excerpt":"Instruction-based video editing has witnessed rapid progress, yet current methods often struggle with precise visual control, as natural language is inherently limited in describing complex visual nuances. Although reference-guided editing offers a robust solution, its potential is currently bottlenecked by the scarcity of high-quality paired training data. To bridge this gap, we introduce a scalable data generation pipeline that transforms existing video editing pairs into high-fidelity training quadruplets, leveraging image generative models to create synthesized reference scaffolds. Using t"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our model achieves significant gains in instruction following and reference fidelity via a progressive multi-stage training curriculum. Extensive experiments demonstrate that our data and architecture establish a new state-of-the-art in controllable video editing.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The image generative models used in the data pipeline produce synthesized reference scaffolds that are high-fidelity and unbiased enough to train a model that generalizes to real user-provided references without introducing artifacts or distribution shifts.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Kiwi-Edit introduces a scalable pipeline to generate RefVIE dataset and a unified model using learnable queries plus reference features to achieve new state-of-the-art in instruction-and-reference guided video editing.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Kiwi-Edit achieves state-of-the-art results in controllable video editing by combining instructions with reference images through a new data pipeline and architecture.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e18ce611c993c105eeb645cb594d0d4ba9bc8bdc9675158d47d432bbde5b9b0d"},"source":{"id":"2603.02175","kind":"arxiv","version":4},"verdict":{"id":"21c21a12-9247-445a-99e0-513e83cb9b5a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T17:32:43.981091Z","strongest_claim":"Our model achieves significant gains in instruction following and reference fidelity via a progressive multi-stage training curriculum. Extensive experiments demonstrate that our data and architecture establish a new state-of-the-art in controllable video editing.","one_line_summary":"Kiwi-Edit introduces a scalable pipeline to generate RefVIE dataset and a unified model using learnable queries plus reference features to achieve new state-of-the-art in instruction-and-reference guided video editing.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The image generative models used in the data pipeline produce synthesized reference scaffolds that are high-fidelity and unbiased enough to train a model that generalizes to real user-provided references without introducing artifacts or distribution shifts.","pith_extraction_headline":"Kiwi-Edit achieves state-of-the-art results in controllable video editing by combining instructions with reference images through a new data pipeline and architecture."},"references":{"count":32,"sample":[{"doi":"","year":null,"title":"- Object identity, attributes (color, shape, material, style), and edit type must be consistent","work_id":"d5d96ab4-c93f-4e63-ae26-2ad1f7cbb1a8","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"- Coherent structure, plausible lighting and texture","work_id":"b6eb58bc-17f3-42e1-a954-5cb7b805697a","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Object not swapped/added, or a completely unrelated object appears","work_id":"1e0d2148-6d9f-4a08-b6ec-661e90d27fa9","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Object is changed, but looks nothing like the reference image (wrong color, shape, or class)","work_id":"3db0ad97-159f-4a41-870d-ed838e14bd14","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Object class is correct, but identity details (texture, specific markings, logos) differ significantly from the reference image","work_id":"0da8ff03-155b-4fe6-a9de-889d62aece5d","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":32,"snapshot_sha256":"4bb9d83768bf47e9c83c3bc6bd7f64d35cbd3aa612d627a0e0fb0bda30271b1e","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"}