Research map / 3D vision / Gaussian splatting
BloomScene: Lightweight Structured 3D Gaussian Splatting for Crossmodal Scene Generation
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 BloomScene, a progressive crossmodal (text/image to 3D) scene generator built on structured anchor-based 3D Gaussian splatting with depth-prior regularization and hash-grid context compression.
Problem
Text- or image-driven 3D scene generation with 3DGS suffers from geometric distortions and artifacts due to photometric-only optimization, and generated scenes require very large storage.
Approach
Combines progressive point-cloud construction with inpainting and monocular depth, adds hierarchical depth-prior losses (Huber-based gradient-aware pixel loss, CMD distribution loss, bilateral smoothness loss), and compresses Scaffold-GS/HAC-style anchors with a structured hash grid providing context for entropy coding and a dynamic quantization scheme.
Key results
- BloomScene reaches 99.22 MB model size versus 571.63 MB for LucidDreamer and 430.55 MB for Invisible-stitch, reported as 5.8x and 4.3x reductions.
- Best CLIP-Score of 31.78 and CLIP-IQA Quality of 0.70 among compared methods, with BRISQUE 20.16 and NIQE 2.92.
- Ablation: removing SCC raises model size from 99.22 MB to 569.33 MB while removing DPR lowers BRISQUE/NIQE performance (BRISQUE 22.15 without DPR).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DPR (ablation) | BloomScene · Multiple scenes (average) | CLIP-Score: 31.38 -> 31.78 | Helped |
| Add SCC compression instead of using original 3DGS (ablation) | BloomScene · Multiple scenes (average) | Size (MB): 569.33 -> 99.22 | Helped |
| Add DPR smoothness loss (ablation) | BloomScene · Multiple scenes (average) | CLIP-Score: 31.68 -> 31.78 | Helped |
| Add DPR distribution loss (ablation) | BloomScene · Multiple scenes (average) | CLIP-Score: 31.72 -> 31.78 | Helped |
| Add DPR pixel loss (ablation) | BloomScene · Multiple scenes (average) | CLIP-Score: 31.54 -> 31.78 | Helped |
Limitations
- Evaluation relies only on 2D reference-free and CLIP-based metrics on 9 prompts, with no 3D geometric metric.
- Depends on pretrained diffusion inpainting and monocular depth estimator, so errors in these propagate into the scene.
Experimental setup
Benchmarks: custom 9-prompt text-to-scene set
Metrics: model size (MB) CLIP-Score CLIP-IQA (Quality, Colorful, Sharp) BRISQUE NIQE
Baselines: Text2Room SceneScape Invisible-stitch WonderJourney LucidDreamer
Builds on: 3D Gaussian Splatting Scaffold-GS HAC LucidDreamer Stable Diffusion
Closest papers by meaning
- Wonderland: Navigating 3D Scenes from a Single Image (CVPR 2025 · similarity 0.90)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.89)
- DiffSplat: Repurposing Image Diffusion Models for Scalable Gaussian Splat Generation (ICLR 2025 · similarity 0.89)
- Scene Splatter: Momentum 3D Scene Generation from Single Image with Video Diffusion Model (CVPR 2025 · similarity 0.89)
- EAP-GS: Efficient Augmentation of Pointcloud for 3D Gaussian Splatting in Few-shot Scene Reconstruction (CVPR 2025 · similarity 0.89)
- DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation (ICLR 2024 · similarity 0.89)
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.