Research map / 3D vision / 3D generation
MIRROR: Make Your Object-Level Multi-View Generation More Consistent with Training-Free Rectification
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 MIRROR, a training-free, plug-and-play rectification method that uses depth-guided trajectory tracking and feature rectification during DDIM sampling to improve multi-view consistency of object-level multi-view diffusion models.
Problem
Object-level multi-view diffusion models produce inconsistent 3D geometry (Janus problem, content drift) and artifacts across views, and retraining to fix this is costly.
Approach
Track pixel-wise correspondences of physical points across adjacent azimuth views using monocular depth from a first-stage sample (Trajectory Tracking Module), then apply a gradient-based feature rectification with dual positive/negative anchors on predicted clean latents during a second DDIM-inverted sampling stage, dropping the UNet Jacobian term for speed.
Key results
- x
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add MIRROR rectification | SyncDreamer · GSO | PSNR: 18.75 -> 19.37 | Helped |
| Add MIRROR rectification | MVD-Fusion · GSO | PSNR: 18.90 -> 20.15 | Helped |
| Add MIRROR rectification | VideoMV (image-based) · GSO | PSNR: 17.64 -> 18.26 | Helped |
| Add MIRROR rectification | VideoMV (text-based) · T3Bench | PSNR: 21.67 -> 24.82 | Helped |
| Use two-stage generation without rectification instead of MIRROR's two-stage rectification (ablation) | VideoMV (text-based) · T3Bench (50 prompts) | PSNR: 23.90 -> 24.82 | Hurt |
| Apply rectification during Stage 1 rather than using the two-stage process (ablation) | VideoMV (text-based) · T3Bench (50 prompts) | PSNR: 12.03 -> 24.82 | Hurt |
| Remove rectification from the baseline single-stage generation (ablation) | VideoMV (text-based) · T3Bench (50 prompts) | PSNR: 21.67 -> 24.82 | Hurt |
| Remove anchor-based feature fusion (ablation) | MIRROR · T3Bench (50 prompts) | PSNR: 24.56 -> 24.82 | Hurt |
| Use only positive anchors instead of dual anchors (ablation) | MIRROR · T3Bench (50 prompts) | PSNR: 23.27 -> 24.82 | Hurt |
| Use only negative anchors instead of dual anchors (ablation) | MIRROR · T3Bench (50 prompts) | PSNR: 24.13 -> 24.82 | Hurt |
Surprising findings
- Applying rectification only in stage 1 (without the second-stage DDIM-inversion resampling) severely degrades results (text-based PSNR drops to 12.03).
Limitations
- Restricted to fixed-elevation circular camera trajectories, limiting arbitrary viewpoint use.
- Limited to object-level generation, not scene-level.
- Cannot correct fundamentally flawed baseline geometry; relies on depth priors that can be inaccurate.
Experimental setup
Datasets: GSO T3Bench
Benchmarks: GSO T3Bench
Metrics: PSNR SSIM LPIPS CLIP-Score inference time
Baselines: SyncDreamer MVD-Fusion VideoMV CoTracker3 epipolar correspondence methods
Builds on: DDIM SyncDreamer VideoMV MVD-Fusion Depth-Anything-V2
Closest papers by meaning
- MirrorVerse: Pushing Diffusion Models to Realistically Reflect the World (CVPR 2025 · similarity 0.89)
- MOVIS: Enhancing Multi-Object Novel View Synthesis for Indoor Scenes (CVPR 2025 · similarity 0.86)
- ViVid-1-to-3: Novel View Synthesis with Video Diffusion Models (CVPR 2024 · similarity 0.85)
- Novel View Synthesis with Pixel-Space Diffusion Models (CVPR 2025 · similarity 0.85)
- DiffusionSfM: Predicting Structure and Motion via Ray Origin and Endpoint Diffusion (CVPR 2025 · similarity 0.85)
- HarmonyView: Harmonizing Consistency and Diversity in One-Image-to-3D (CVPR 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.