Research map / 3D vision / NeRF and radiance fields
NeRF Analogies: Example-Based Visual Attribute Transfer for NeRFs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Transfers appearance from a source NeRF onto a target NeRF's geometry via DINO ViT feature correspondence, training a new NeRF analogy with direct supervised learning on point-appearance pairs.
Problem
Transferring semantically meaningful appearance from a source NeRF onto arbitrary target 3D geometry while keeping the result multi-view consistent, a gap left by 2D image analogies and stylization methods that are not 3D-consistent.
Approach
Generalizes image analogies to NeRFs by computing semantic affinity between renderings of target geometry and source NeRF renderings using DINO ViT features, mapping per-point appearance, and directly supervising a view-dependent NeRF (no differentiable rendering needed) with an added difference-of-Gaussians edge loss.
Key results
- Ours achieves BPSNR 36.16 and BSSIM .984 vs. 33.06 / .968 for Deep Image Analogies, the strongest baseline.
- In the user study (42 participants), 68.4% combined preference for the proposed method vs. 23.0% for DIA.
- Ours is preferred in transfer quality by 58.5% vs 28.6% for DIA and CLIP direction consistency is 0.992 (highest).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use NeRF Analogies instead of Deep Image Analogies | NeRF appearance transfer · Object-pair evaluation | BPSNR: 33.06 -> 36.16 | Helped |
| use NeRF Analogies instead of Deep Image Analogies | NeRF appearance transfer · Object-pair evaluation | BSSIM: .968 -> .984 | Helped |
| use NeRF Analogies instead of Deep Image Analogies | NeRF appearance transfer · Object-pair evaluation | CLIP: .983 -> .992 | Helped |
| use NeRF Analogies instead of Deep Image Analogies | NeRF appearance transfer · User study | Multi-view consistency preference: 20.5% -> 76.7% | Helped |
| add edge loss (ablation) | NeRF Analogies · Object-pair evaluation | Appearance detail | Helped |
| retain target texture (ablation) | NeRF Analogies · Object-pair evaluation | Semantic appearance transfer | Helped |
Surprising findings
- Deep Image Analogies performs surprisingly well on most metrics despite lacking 3D information, attributed to its patch-based output being drawn from 3D-consistent input.
Limitations
- Point-based appearance transfer cannot transfer texture.
- DINO correspondences are inaccurate on specular or rotation-symmetric objects, leading to color errors in viewing directions.
- Requires roughly aligned objects and textured target geometry for good DINO feature performance.
Experimental setup
Datasets: MiP-NeRF 360 Tanks and Temples
Benchmarks: MiP-NeRF 360 Tanks and Temples
Metrics: BPSNR BSSIM CLIP direction consistency user preference
Baselines: Neural Style Transfer WCT Deep Image Analogies SNeRF Instruct-NeRF2NeRF
Builds on: image analogies DINO-ViT Instant-NGP deep image analogies SNeRF
Closest papers by meaning
- NeRFDeformer: NeRF Transformation from a Single View via 3D Scene Flows (CVPR 2024 · similarity 0.90)
- To View Transform or Not to View Transform: NeRF-based Pre-training Perspective (ICLR 2026 · similarity 0.90)
- Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields (CVPR 2022 · similarity 0.90)
- Aug-NeRF: Training Stronger Neural Radiance Fields With Triple-Level Physically-Grounded Augmentations (CVPR 2022 · similarity 0.89)
- GP-NeRF: Generalized Perception NeRF for Context-Aware 3D Scene Understanding (CVPR 2024 · similarity 0.88)
- Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates (CVPR 2024 · similarity 0.88)
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.