Research map / 3D vision / Gaussian splatting
MiraGe: Editable 2D Images using 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 MiraGe, which encodes a 2D image as flat 3D Gaussians (GaMeS triangle soup) placed on a plane with a mirror camera, enabling 2D/3D editing and physics-engine animation while outperforming GaussianImage on Kodak and DIV2K reconstruction.
Problem
Existing Gaussian-based 2D image representations like GaussianImage achieve good compression and reconstruction but do not support user-driven editing of image content, unlike 3DGS scenes.
Approach
Represent a 2D image as flat 3D Gaussians parameterized by GaMeS (each Gaussian as a triangle), placed on the XZ plane with a mirrored second camera for augmentation, so the image can be edited, moved in 3D, and animated with physics engines (Taichi_elements, Blender).
Key results
- MiraGe-100k reaches 59.52 PSNR / 0.9999 MS-SSIM on Kodak versus 44.08 PSNR for GaussianImage-70k.
- On DIV2K, MiraGe-100k achieves 54.54 PSNR versus 41.48 for GaussianImage-100k (39.53 for GaussianImage-70k).
- Ablation shows the mirror camera augmentation improves Amorphous PSNR on Kodak from 51.56 to 59.52 at 100k initial Gaussians, at the cost of longer training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MiraGe-100k instead of GaussianImage-70k | 2D image reconstruction · Kodak | PSNR: 44.08 -> 59.52 | Helped |
| Use MiraGe-100k instead of GaussianImage-100k* | 2D image reconstruction · DIV2K | PSNR: 41.48 -> 54.54 | Helped |
| Add a mirror camera during training (ablation) | Amorphous MiraGe, 100k Gaussians · Kodak | PSNR: 51.56 -> 59.52 | Helped |
| Add a mirror camera during training (ablation) | Amorphous MiraGe, 100k Gaussians · DIV2K | PSNR: 46.00 -> 54.54 | Helped |
| Add a mirror camera during training (ablation) | Amorphous MiraGe, 100k Gaussians · Kodak | training time: 448.73 -> 639.66 | Hurt |
| Add a mirror camera during training (ablation) | Amorphous MiraGe, 100k Gaussians · Kodak | FPS: 583.28 -> 620.10 | Mixed |
| Increase initialization from 100k to 150k Gaussians (ablation) | Amorphous MiraGe · Kodak | PSNR: 59.52 -> 62.12 | Helped |
| Increase initialization from 100k to 150k Gaussians (ablation) | Amorphous MiraGe · DIV2K | PSNR: 54.54 -> 56.40 | Helped |
| Restrict Gaussian movement to the XZ plane instead of allowing free 3D movement (ablation) | MiraGe · Kodak | PSNR: 59.52 -> 48.82 | Hurt |
| Restrict Gaussian movement to the XZ plane instead of allowing free 3D movement (ablation) | MiraGe · DIV2K | PSNR: 54.54 -> 46.32 | Hurt |
Surprising findings
- Graphite and 2D variants without mirror camera lose substantial reconstruction quality relative to Amorphous, showing sensitivity to Gaussian control method.
Limitations
- Model is not generative, so moving foreground elements leaves gaps in the background that require inpainting.
- Requires encoding more parameters than GaussianImage, reducing compression efficiency.
- Significant modifications can introduce visual artifacts.
Experimental setup
Datasets: Kodak DIV2K
Benchmarks: Kodak DIV2K
Metrics: PSNR MS-SSIM LPIPS FPS training time
Baselines: GaussianImage SIREN WIRE I-NGP NeuRBF 3DGS DragGAN PhysGen
Builds on: GaussianImage 3D Gaussian Splatting GaMeS MPM
Closest papers by meaning
- Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction (CVPR 2024 · similarity 0.87)
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.87)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.87)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.87)
- Gaussian Shadow Casting for Neural Characters (CVPR 2024 · similarity 0.87)
- Mani-GS: Gaussian Splatting Manipulation with Triangular Mesh (CVPR 2025 · similarity 0.87)
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.