Research map / 3D vision / 3D generation
Direct 3D-Aware Object Insertion via Decomposed Visual Proxies
Jingbo Gong, Yikai Wang, Yushi Lan, Yuhao Wan, Ziheng Ouyang, Rui Zhao, Ming-Ming Cheng, Qibin Hou, Chen Change Loy
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 hybrid real-world training data (ablation) | DIRECT FLUX base · Ablation benchmark | CLIP-I / Matching Error: CLIP-I 0.904 -> 0.943; Matching Error 26.9 -> 22.7 | Improved |
| add context guidance (ablation) | DIRECT FLUX · Ablation benchmark | PSNR: 22.56 -> 22.74 | Improved |
| add context guidance (ablation) | DIRECT FLUX · Ablation benchmark | Matching Error: 22.7 -> 20.7 | Improved |
| add shape-decomposed mask augmentation (ablation) | DIRECT FLUX · Ablation benchmark | LPIPS / Matching Error: LPIPS 0.190 -> 0.155; Matching Error 20.7 -> 19.0 | Improved |
| add progressive resolution training (ablation) | DIRECT FLUX · Ablation benchmark | PSNR / Matching Error: PSNR 22.89 -> 23.09; Matching Error 19.0 -> 17.8 | Improved |
| use RGB geometry guidance instead of surface normal maps (ablation) | DIRECT · Symmetric road sign example | orientation correctness (qualitative) | Improved |
| use decomposed injection instead of naive concatenation of guidance signals (ablation) | DIRECT · Textured proxy artifacts (qualitative) | feature entanglement / texture sharpness | Improved |
| switch backbone from SD to FLUX | DIRECT · Object insertion benchmark | PSNR / LPIPS: PSNR 21.66 -> 23.09; LPIPS 0.206 -> 0.147 | Improved |
| compare against TRELLIS + ZeroComp intrinsic-guided compositing | DIRECT (FLUX) · Object insertion benchmark | Matching Error: 5.2 vs 17.8 | Hurt |
| increase relative rotation between input and target pose (135-180 vs 0-45 degrees) | DIRECT (FLUX) · Pose-change bins | Matching Error: 18.1 -> 15.7 | No clear effect |
| use Ours instead of Object3DIT in SD pipeline | SD-based insertion · Object insertion benchmark | Matching Error: 135.7 -> 21.4 | Improved |
| add 3D proxy generation overhead versus Object3DIT/TRELLIS | DIRECT (SD) · Latency | Overall latency (s): 8.610 / 11.500 -> 9.542 | Mixed |
Closest papers by meaning
- Thinking in Boxes: 3D Editing in Real Images Made Easy (NeurIPS 2026 · similarity 0.87)
- InstaInpaint: Instant 3D-Scene Inpainting with Masked Large Reconstruction Model (NeurIPS 2025 · similarity 0.86)
- ElasticFit: Fit-Aware 3D Object Insertion via VLM Reasoning and Generative Adaptation (NeurIPS 2026 · similarity 0.86)
- HOComp: Interaction-Aware Human-Object Composition (NeurIPS 2025 · similarity 0.86)
- MVInpainter: Learning Multi-View Consistent Inpainting to Bridge 2D and 3D Editing (NeurIPS 2024 · similarity 0.86)
- Neural Assets: 3D-Aware Multi-Object Scene Synthesis with Image Diffusion Models (NeurIPS 2024 · 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.
