Research map / 3D vision / Gaussian splatting
DiET-GS: Diffusion Prior and Event Stream-Assisted Motion Deblurring 3D Gaussian Splatting
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 DiET-GS, a two-stage event-and-diffusion-prior method that builds motion-deblurring 3D Gaussian Splatting using EDI constraints from event streams, then refines edges with a renoised score distillation stage (DiET-GS++).
Problem
Reconstructing sharp 3D representations from motion-blurred multi-view images, where prior methods either get inaccurate colors or lose fine-grained edge details.
Approach
Combines event double integral (EDI) constraints in both RGB and a learnable camera-response-based brightness domain with event reconstruction loss for 3DGS, then adds a second stage that attaches learnable per-Gaussian latent features rendered as a residual and optimized with renoised score distillation from a pretrained diffusion model.
Key results
- On EvDeblur-Blender, DiET-GS reaches 26.69 PSNR and 0.8607 SSIM, versus 24.76 PSNR for Ev-DeblurNeRF.
- On EvDeblur-CDAVIS, DiET-GS reaches 34.22 PSNR and 0.9223 SSIM, versus 32.30 PSNR for Ev-DeblurNeRF.
- DiET-GS++ achieves the best MUSIQ and CLIP-IQA (e.g., 59.91 MUSIQ on Blender, 50.44 on CDAVIS) at a small cost in PSNR and SSIM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiET-GS rather than the strongest baseline | EvDeblur-blender · EvDeblur-blender | PSNR: 24.76 -> 26.69 | Helped |
| Use DiET-GS rather than the strongest baseline | EvDeblur-CDAVIS · EvDeblur-CDAVIS | PSNR: 32.95 -> 34.22 | Helped |
| Add Stage 2 diffusion refinement to DiET-GS | DiET-GS · EvDeblur-blender | MUSIQ: 57.67 -> 59.91 | Helped |
| Add Stage 2 diffusion refinement to DiET-GS | DiET-GS · EvDeblur-CDAVIS | PSNR: 34.22 -> 33.16 | Hurt |
| Add event reconstruction loss Lev (ablation) | DiET-GS · Figures sample | PSNR: 29.73 -> 32.74 | Helped |
| Add grayscale EDI loss Ledi gray (ablation) | DiET-GS · Figures sample | PSNR: 32.74 -> 33.91 | Helped |
| Use color EDI loss instead of grayscale EDI loss (ablation) | DiET-GS · Figures sample | PSNR: 33.91 -> 34.48 | Helped |
| Add color EDI loss alongside grayscale EDI loss (ablation) | DiET-GS · Figures sample | PSNR: 33.91 -> 34.92 | Helped |
| Add EDI simulation loss Ledi simul (ablation) | DiET-GS · Figures sample | PSNR: 34.92 -> 35.04 | Helped |
| Add Stage 1 RSD diffusion loss Lrsd (ablation) | DiET-GS · Figures sample | MUSIQ: +0.33 | Helped |
| Add Stage 1 RSD diffusion loss Lrsd (ablation) | DiET-GS · Figures sample | PSNR: 35.04 -> 34.89 | Hurt |
| Add Stage 2 RSD with learnable Gaussian features (ablation) | DiET-GS · Figures sample | MUSIQ: 45.37 -> 51.71 | Helped |
Surprising findings
- DiET-GS++ improves no-reference quality (MUSIQ, CLIP-IQA) while slightly lowering PSNR and SSIM relative to DiET-GS.
- Using EDI-processed sharp images as the diffusion conditioning input degraded results due to artifacts, so ground-truth blurry images were used instead.
Limitations
- Joint optimization of event losses and RSD loss weakens the diffusion prior's effect, requiring a second stage.
- DiET-GS++ trades slight PSNR/SSIM drop for perceptual quality because outputs can deviate from ground truth.
Experimental setup
Datasets: EvDeblur-Blender EvDeblur-CDAVIS
Benchmarks: EvDeblur-Blender EvDeblur-CDAVIS
Metrics: PSNR SSIM LPIPS MUSIQ CLIP-IQA
Baselines: MPRNet EDI+GS EFNet+GS BAD-NeRF BAD-GS E2NeRF Ev-DeblurNeRF
Builds on: 3D Gaussian Splatting Ev-DeblurNeRF DiSR-NeRF EDI Renoised Score Distillation
Closest papers by meaning
- GS-Blur: A 3D Scene-Based Dataset for Realistic Image Deblurring (NeurIPS 2024 · similarity 0.90)
- EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- EvGGS: A Collaborative Learning Framework for Event-based Generalizable Gaussian Splatting (ICML 2024 · similarity 0.89)
- BARD-GS: Blur-Aware Reconstruction of Dynamic Scenes via Gaussian Splatting (CVPR 2025 · similarity 0.89)
- GS2E: Gaussian Splatting is an Effective Data Generator for Event Stream Generation (NeurIPS 2025 · similarity 0.89)
- HiCoM: Hierarchical Coherent Motion for Dynamic Streamable Scenes with 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.88)
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.