CasDeblurGS: Cascaded 2D-to-3D Multi-view Consistency for 3D Gaussian Splatting from Two Blurry Images

Haeyun Choi1*, MinHyuk Jang2*, I-Gil Kim2

1 University of Virginia, 2 KT R&D Center

* Equal contribution

ECCVW 2026 (MUSTCV Workshop)

Paper Coming Soon Supplementary Coming Soon Code Coming Soon BibTeX
Qualitative comparison of Inputs, Ours, CoherentGS, Difix3D+, and GAURA
From only two blurry input views, our framework produces sharper details and more coherent novel views than strong sparse-view blurry reconstruction baselines.

TL;DR

CasDeblurGS reconstructs a coherent 3D representation from only two blurry input views by cascading reliable local 2D guidance with global 3D re-render guidance, enabling high-quality novel view synthesis.

Method

CasDeblurGS progressively converts two blurry, pose-free inputs into reliable local 2D and global 3D guidance. Occlusion-aware cross-view correspondence first produces consistent intermediate restorations, which are then lifted into a provisional 3D representation and re-rendered for final restoration. The refined views are finally reconstructed into coherent 3D Gaussians for novel-view synthesis.

Overview of the CasDeblurGS method
  1. Pose-free InputsTwo blurry images and known intrinsics are used without pre-estimated camera poses.
  2. Local 2D GuidanceOCGM stabilizes the views, validates bidirectional RAFT flow, and forms masked cross-view warps for reliable deblurring.
  3. Global 3D GuidanceA frozen pose-free 3DGS backbone builds a provisional representation and re-renders dense scene-level guidance for final restoration.
  4. Final 3D RepresentationThe final restored views are reconstructed into coherent 3D Gaussians for novel-view synthesis, without test-time optimization.

Quantitative Results

Quantitative comparison table on real-world and synthetic datasets using PSNR, SSIM, and LPIPS

Qualitative Results

Qualitative comparison of Inputs, SE-GS, GAURA, CoherentGS, DAVANet, Difix3D+, Ours, and ground truth

BibTeX

@inproceedings{choi2026casdeblurgs,
  title     = {CasDeblurGS: Cascaded 2D-to-3D Multi-view Consistency for 3D Gaussian Splatting from Two Blurry Images},
  author    = {Choi, Haeyun and Jang, MinHyuk and Kim, I-Gil},
  booktitle = {ECCV Workshop on Computer Vision for Multimedia Spatial Intelligence through Time},
  year      = {2026}
}