Research map / 3D vision / 3D generation

3DEnhancer: Consistent Multi-View Diffusion for 3D Enhancement

CVPR 2025 · 3D vision › 3D generation · approach: Score distillation for text-to-3D · 5 citations (OpenAlex)

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 3DEnhancer, a multi-view latent diffusion (PixArt-Σ DiT) enhancer with row attention and near-view epipolar aggregation that refines coarse multi-view images and 3D reconstructions from Objaverse-trained degradation pairs.

Problem

Multi-view diffusion and feed-forward 3D generation yield low-resolution, view-inconsistent multi-view images, limiting 3D quality; existing 2D/video enhancers fail to keep multi-view consistency.

Approach

Adapt a text-to-image DiT (PixArt-Σ) into a pose-aware multi-view enhancer using efficient multi-view row attention plus a trainable near-view epipolar feature aggregation module, trained with degradation augmentations and noise-level control.

Key results

• On the Objaverse synthetic set, 3DEnhancer reaches 27.53 PSNR / 0.9265 SSIM / 0.0626 LPIPS, beating the best baseline RealBasicVSR at 26.21 PSNR.
• On the in-the-wild multi-view enhancement set, 3DEnhancer achieves the best FID (108.40) and IS (7.93), while RealBasicVSR has higher MUSIQ (74.07 vs 73.32).
• For 3D reconstruction enhancement (LGM outputs), 3DEnhancer gets FID 71.78 vs 72.98 for StableSR, the best baseline.

Measured effects reported in the paper

Surprising findings

• RealBasicVSR achieves higher MUSIQ than 3DEnhancer on the in-the-wild set despite producing visually implausible images.

Limitations

• Relies on the assumption that cameras have near-horizontal viewpoints (low elevation) for the row-attention epipolar approximation.
• Baselines fine-tuned on Objaverse may not represent their best-possible settings; evaluation mostly on synthetic data.

Experimental setup

Datasets: Objaverse G-buffer Objaverse GSO

Benchmarks: Objaverse synthetic test set (500 objects) in-the-wild test set (400 instances)

Metrics: PSNR SSIM LPIPS FID Inception Score MUSIQ

Baselines: Real-ESRGAN StableSR RealBasicVSR Upscale-A-Video

Builds on: PixArt-Σ ControlNet TokenFlow LGM StableSR

Closest papers by meaning

• Era3D: High-Resolution Multiview Diffusion using Efficient Row-wise Attention (NeurIPS 2024 · similarity 0.88)
• Diffusion$^2$: Dynamic 3D Content Generation via Score Composition of Video and Multi-view Diffusion Models (ICLR 2025 · similarity 0.87)
• Cycle3D: High-quality and Consistent Image-to-3D Generation via Generation-Reconstruction Cycle (AAAI 2025 · similarity 0.87)
• Ouroboros3D: Image-to-3D Generation via 3D-aware Recursive Diffusion (CVPR 2025 · similarity 0.87)
• In-N-Out: Lifting 2D Diffusion Prior for 3D Object Removal via Tuning-Free Latents Alignment (NeurIPS 2024 · similarity 0.87)
• Sharp-It: A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation (CVPR 2025 · similarity 0.86)