{"paper":{"title":"DEVIS-GRPO: Unleashing GRPO on Dynamic Extreme View Synthesis","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Accumulating small camera increments during sampling lets a policy-gradient model handle extreme-view video generation without paired large-motion training data.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Fang Liu, Huimin Wu, Licheng Jiao, Lingling Li, Qing Li, Yi Zuo","submitted_at":"2026-05-16T11:14:18Z","abstract_excerpt":"Trajectory-controlled video generation has become essential for controllable video generation. While current methods perform well under small-view camera motions, they degrade significantly with large-view motions. Existing solutions for extreme-view synthesis typically require dedicated video pairs, demanding substantial annotation effort. To address these limitations, we propose Dynamic Extreme VIew Synthesis-GRPO (DEVIS-GRPO), a GRPO-based framework for trajectory-controlled video generation, the first online policy gradient method for extreme view video generation. Central to our approach "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"we propose Dynamic Extreme VIew Synthesis-GRPO (DEVIS-GRPO), a GRPO-based framework for trajectory-controlled video generation, the first online policy gradient method for extreme view video generation. Central to our approach is a novel sampling strategy: Accumulative Dynamic Extreme VIew Synthesis (ADEVIS), which achieves large-view camera motions by progressively accumulating small-view increments.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"Progressively accumulating small-view increments reliably produces high-quality large-view motions and increased sampling diversity without introducing artifacts that the multi-level reward cannot filter, eliminating the need for expensive paired large-view videos.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DEVIS-GRPO applies online policy gradients with an accumulative small-to-large view sampling strategy and multi-level rewards to improve trajectory-controlled extreme view video generation, reporting gains on Kubric-4D, iPhone, and DL3DV datasets.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Accumulating small camera increments during sampling lets a policy-gradient model handle extreme-view video generation without paired large-motion training data.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d3d57330badd617e3885d0355c4d0a09b483071150c08be4944325541650e26f"},"source":{"id":"2605.16937","kind":"arxiv","version":1},"verdict":{"id":"3124d67b-a514-422e-a945-2d730636bdbc","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T20:50:48.173269Z","strongest_claim":"we propose Dynamic Extreme VIew Synthesis-GRPO (DEVIS-GRPO), a GRPO-based framework for trajectory-controlled video generation, the first online policy gradient method for extreme view video generation. Central to our approach is a novel sampling strategy: Accumulative Dynamic Extreme VIew Synthesis (ADEVIS), which achieves large-view camera motions by progressively accumulating small-view increments.","one_line_summary":"DEVIS-GRPO applies online policy gradients with an accumulative small-to-large view sampling strategy and multi-level rewards to improve trajectory-controlled extreme view video generation, reporting gains on Kubric-4D, iPhone, and DL3DV datasets.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"Progressively accumulating small-view increments reliably produces high-quality large-view motions and increased sampling diversity without introducing artifacts that the multi-level reward cannot filter, eliminating the need for expensive paired large-view videos.","pith_extraction_headline":"Accumulating small camera increments during sampling lets a policy-gradient model handle extreme-view video generation without paired large-motion training data."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.16937/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T21:01:19.133724Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T21:00:54.351634Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"cited_work_retraction","ran_at":"2026-05-19T20:21:57.751616Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T18:41:56.249228Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T18:33:26.331368Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"c8b9238b24030c0644a7a437b1689948db13854e7c3e03fc09e36ad9cb5b9843"},"references":{"count":79,"sample":[{"doi":"","year":null,"title":"Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =","work_id":"b4323590-3548-479f-9bdb-84b233131806","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Cami2v: Camera-controlled image-to-video diffu- sion model","work_id":"dfc5b040-8e44-47be-8714-65cc5b5f8400","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Lindell and Sergey Tulyakov , booktitle =","work_id":"3da69074-d191-45c2-ae9e-b478ef8800c7","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"CameraCtrl: Enabling Camera Control for Text-to-Video Generation","work_id":"1c05c278-c023-4ef0-a359-25a41f1065eb","ref_index":4,"cited_arxiv_id":"2404.02101","is_internal_anchor":true},{"doi":"","year":null,"title":"Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages =","work_id":"735560ec-d143-465b-885e-689a29b5dd34","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":79,"snapshot_sha256":"db1f2730fd84bf5ebdf9e75a53fd8286440fccef68866355081898f7228af8ac","internal_anchors":14},"formal_canon":{"evidence_count":2,"snapshot_sha256":"eadbb74667398b7eb4f92594ff7ebaa82c1cf2dfe6ca686753956c12f7b71bc3"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}