GeoRect4D: Geometry-Compatible Generative Rectification for Dynamic Sparse-View 3D Reconstruction

2026-04-22 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

GeoRect4D is a new unified framework designed for sparse-view dynamic 3D reconstruction, addressing common issues like geometric collapse, trajectory drift, and floating artifacts that arise from the ill-posed nature of reconstructing dynamic 3D scenes from sparse multi-view videos. Unlike prior methods that integrate generative priors, GeoRect4D couples explicit 3D consistency with generative refinement through a closed-loop optimization. It features a degradation-aware feedback mechanism, combining a robust anchor-based dynamic 3D Gaussian Splatting (3DGS) substrate with a single-step diffusion rectifier. This rectifier employs a structural locking mechanism and spatiotemporal coordinated attention to maintain physical plausibility while restoring missing content. Additionally, GeoRect4D uses a progressive optimization strategy involving stochastic geometric purification to remove floaters and generative distillation to enhance texture details in the explicit representation, achieving superior performance in fidelity, perceptual quality, and spatiotemporal consistency.

Key takeaway

For research scientists working on dynamic 3D reconstruction from sparse multi-view videos, GeoRect4D offers a robust approach to mitigate common artifacts like geometric collapse and floating. You should consider its integrated explicit 3D consistency and generative refinement strategy to improve reconstruction fidelity and temporal consistency in your models, especially when dealing with limited input data.

Key insights

GeoRect4D integrates explicit 3D consistency with generative refinement to reconstruct dynamic 3D scenes from sparse views.

Principles

• Couple 3D consistency with generative priors.
• Use degradation-aware feedback for refinement.
• Employ structural locking for physical plausibility.

Method

GeoRect4D uses a closed-loop optimization with a dynamic 3DGS substrate and a diffusion rectifier with structural locking. It applies progressive optimization via geometric purification and generative distillation.

In practice

• Apply 3DGS for dynamic scene representation.
• Integrate diffusion models for detail hallucination.
• Utilize spatiotemporal attention for consistency.

Topics

• Dynamic 3D Reconstruction
• Sparse-View Videos
• Generative Rectification
• 3D Gaussian Splatting
• Diffusion Models

Best for: Research Scientist, AI Scientist, Computer Vision Engineer

Related on AIssential

• Rein3D: Reinforced 3D Indoor Scene Generation with Panoramic Video Diffusion Models — Rein3D integrates 3D Gaussian Splatting with panoramic video diffusion to restore and refine 3D indoor scenes from sparse inputs.
• FlowObject: Flow Steering for Bridging Generative Priors and Reconstruction Fidelity — FlowObject bridges generative priors and observational fidelity for robust 3D object reconstruction from sparse views.
• ReCoSplat: Autoregressive Feed-Forward Gaussian Splatting Using Render-and-Compare — ReCoSplat uses a Render-and-Compare module and KV cache compression for robust, online novel view synthesis.
• Feed-Forward 3D Scene Modeling: A Problem-Driven Perspective — Feed-forward 3D reconstruction offers efficient, generalizable scene modeling by directly mapping 2D inputs to 3D representations.
• Stream3D: Sequential Multi-View 3D Generation via Evidential Memory — Stream3D uses an evidential memory and evidence score mechanism to enable consistent 3D generation from sequential multi-view inputs without retraining existing models.
• Real2SAM2Real: Generative 3D Caches as Complementary Context for Video Diffusion — Real2SAM2Real uses editable 3D caches from lifting models to provide robust geometric scaffolds for Video Diffusion Models, improving control and consistency.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.