Research map / 3D vision / NeRF and radiance fields
Is Vanilla MLP in Neural Radiance Field Enough for Few-shot View Synthesis?
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.
Proposes mi-MLP, a NeRF network-structure change that feeds location and view-direction inputs into every MLP layer, plus separate color/density branches and two regularizers, to reduce overfitting in few-shot view synthesis.
Problem
NeRF overfits to the few known input views in few-shot novel view synthesis, producing geometry on 2D planes and floating/background artifacts; prior work addressed this with learned priors or extra regularization but kept the vanilla MLP structure.
Approach
Address few-shot overfitting from the network structure: multi-input MLP (mi-MLP) concatenates encoded position and view direction into each layer, shortening input-output paths and keeping capacity; colors and density are modeled by separate branches with different positional-encoding frequencies, plus a background regularizer and sampling annealing.
Key results
- mi-MLP boosts the vanilla NeRF baseline PSNR on Blender 8-view from 14.73 to 24.23 (ablation, 4-view table shows 20.38 for full method).
- Full method reaches 20.38 PSNR (4-view) and 24.70 PSNR (8-view) on Blender, state of the art among compared methods.
- On LLFF 3-view the method reaches 19.75 PSNR, and on Shiny 3-view 18.25 PSNR, outperforming FreeNeRF (17.20).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the vanilla MLP with per-layer inputs incorporation (mi-MLP) (ablation) | Vanilla-NeRF · Blender, 8 views | PSNR: 14.73 -> 24.12 | Helped |
| Replace the vanilla MLP with per-layer inputs incorporation (mi-MLP) (ablation) | Vanilla-NeRF · LLFF, 3 views | PSNR: 13.34 -> 18.12 | Helped |
| Add separate color/density modeling to mi-MLP (ablation) | mi-MLP · Blender, 8 views | PSNR: 24.12 -> 24.23 | Helped |
| Add separate color/density modeling to mi-MLP (ablation) | mi-MLP · LLFF, 3 views | PSNR: 18.12 -> 16.87 | Hurt |
| Add background regularization to mi-MLP (ablation) | mi-MLP · Blender, 8 views | PSNR: 24.12 -> 24.42 | Helped |
| Add sampling annealing to mi-MLP (ablation) | mi-MLP · LLFF, 3 views | PSNR: 18.12 -> 18.88 | Helped |
| Combine separate color/density modeling and background regularization with mi-MLP (ablation) | mi-MLP · Blender, 8 views | PSNR: 24.12 -> 24.70 | Helped |
| Combine separate color/density modeling and sampling annealing with mi-MLP (ablation) | mi-MLP · LLFF, 3 views | PSNR: 18.12 -> 19.75 | Helped |
| Use the proposed method instead of FreeNeRF | Few-shot view synthesis · LLFF, 6 views | PSNR: 23.73 -> 23.57 | Hurt |
Surprising findings
- Simply reducing NeRF parameters alleviates overfitting but loses detail; mi-MLP keeps capacity while still avoiding overfitting.
- Regularization-based methods such as DietNeRF and RegNeRF perform worse than vanilla NeRF on the Shiny dataset due to mismatch with view-dependent effects.
Limitations
- Rendered views still show noticeable artifacts when using mi-MLP alone, requiring separate color/density modeling and regularizers.
- Results on LLFF at 6 views do not beat all baselines in PSNR, attributed to differing baseline choices (vanilla NeRF vs Mip-NeRF).
Experimental setup
Datasets: Blender LLFF Shiny
Benchmarks: Blender LLFF Shiny
Metrics: PSNR SSIM LPIPS geometric average
Baselines: Vanilla-NeRF Mip-NeRF Ref-NeRF DietNeRF InfoNeRF RegNeRF MixNeRF FreeNeRF MVSNeRF PixelNeRF SRF
Builds on: NeRF positional encoding FreeNeRF DietNeRF RegNeRF
Closest papers by meaning
- Few-Shot Neural Radiance Fields under Unconstrained Illumination (AAAI 2024 · similarity 0.91)
- NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · similarity 0.90)
- Bright-NeRF: Brightening Neural Radiance Field with Color Restoration from Low-Light RAW Images (AAAI 2025 · similarity 0.89)
- MuRF: Multi-Baseline Radiance Fields (CVPR 2024 · similarity 0.89)
- Enhancing Neural Radiance Fields with Adaptive Multi-Exposure Fusion: A Bilevel Optimization Approach for Novel View Synthesis (AAAI 2024 · similarity 0.89)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (NeurIPS 2024 · similarity 0.89)
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.