Research map / Video / Video generation
InsertAnywhere: Geometrically Grounded and Optics-Aware Video Object Insertion
Hoiyeong Jin, Hyojin Jang, Junha Hyung, Jeongho Kim, Kinam Kim, Dongjin Kim, Huijin choi, Hyeonji Kim, Choo Jaegul
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Optics-Aware Representation Alignment loss (ablation) | ROSE++ LoRA VACE · ROSE++ held-out (25 samples), Optics mask | PSNR: 19.9065 -> 27.2362 | Improved |
| add alignment loss (ablation) | ROSE++ LoRA VACE · ROSE++ held-out (25 samples), Object mask | PSNR: 26.0282 -> 25.8785 | No clear effect |
| add alignment loss (ablation) | ROSE++ LoRA VACE · ROSE++ held-out (25 samples), unedited region | PSNR: 33.31 -> 33.30 | No clear effect |
| fine-tune VACE on ROSE++ with 4D mask (ablation) | VACE (our mask) · VOIBench | CLIP-I: 0.7368 -> 0.7485 | Improved |
| use coarse bbox mask instead of 4D mask when fine-tuning on ROSE++ (ablation) | VACE ROSE++ · VOIBench | CLIP-I: 0.7485 -> 0.6383 | Hurt |
| add first-frame anchoring (ablation) | 4D mask VACE · VOIBench | DINO-I: 0.3861 -> 0.5135 | Improved |
| add first-frame anchoring (ablation) | 4D mask VACE · VOIBench | Background Consistency: 0.9232 -> 0.9175 | Hurt |
| add ROSE++ LoRA fine-tuning (ablation) | anchored VACE · VOIBench | Imaging Quality: 0.6318 -> 0.7101 | Improved |
| add alignment loss on top of LoRA (ablation) | Config (c) · VOIBench | Imaging Quality: 0.7101 -> 0.7473 | Improved |
| add alignment loss on top of LoRA (ablation) | Config (c) · VOIBench | DINO-I: 0.5678 -> 0.5669 | No clear effect |
| use ROSE++ generated references instead of random-frame crops (ablation) | Ours · VOIBench | Multi-View Consistency: 0.5295 -> 0.5865 | Improved |
| 4D-aware mask plus optics-aware pipeline vs Kling | InsertAnywhere · VOIBench | CLIP-I: 0.6349 -> 0.8132 | Improved |
Closest papers by meaning
- Direct 3D-Aware Object Insertion via Decomposed Visual Proxies (ICML 2026 · similarity 0.89)
- Instruct 4D-to-4D: Editing 4D Scenes as Pseudo-3D Scenes Using 2D Diffusion (CVPR 2024 · similarity 0.86)
- MV-CoLight: Efficient Object Compositing with Consistent Lighting and Shadow Generation (NeurIPS 2025 · similarity 0.86)
- ElasticFit: Fit-Aware 3D Object Insertion via VLM Reasoning and Generative Adaptation (NeurIPS 2026 · similarity 0.85)
- PISCO: Precise Video Instance Insertion with Sparse Control (NeurIPS 2026 · similarity 0.85)
- InstaInpaint: Instant 3D-Scene Inpainting with Masked Large Reconstruction Model (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
