A paper- and code-grounded roadmap with twelve deduplicated ideas, recent-literature novelty checks, reviewer-risk analysis, a refined top proposal, and a claim-driven experiment plan. No GPU experiment was run.
- 3D Gaussian Splatting
- GaussianGPT
- Autoregressive generation
- Research ideas
- Experiment planning
A code-grounded survey of the latest 12 months of 3D Gaussian Splatting research, with ten falsifiable topics, novelty checks, reviewer risks, GPU budgets, a refined top proposal, and a claim-driven experiment plan.
- 3D Gaussian Splatting
- ReSplat
- Research ideas
- Novelty check
- Experiment planning
A technical reading of Scal3R, covering global context memory, test-time adaptation, cross-chunk synchronization, long-sequence experiments, and reproducibility boundaries.
- 3D reconstruction
- Test-time training
- Global context
- Visual computing
Read analysis An academic reading of visibility-aware TSDF masking, sparse VAE compression, flow matching, ControlNet completion, experiments, source code, and reproducibility boundaries.
- 3D scene completion
- Flow matching
- TSDF
- Generative models
Read analysis An academic reading of GaussianGPT, connecting sparse LFQ compression, position-feature tokenization, 3D RoPE, causal sampling, scene completion, outpainting, source code, and reproducibility limits.
- 3D Gaussian Splatting
- Autoregressive generation
- Scene synthesis
- Transformers
A technical reading of WorldMesh, covering geometry-first layout construction, object reconstruction, progressive wall-texture accumulation, depth validation, geometry-regularized 3DGS, and reproducibility limits.
- 3D scene generation
- Mesh conditioning
- Image diffusion
- Gaussian splatting
A complete reading of compact Gaussian initialization, rendering-error feedback, recurrent 3D updates, generalization, ablations, appendices, and the official implementation.
- 3D Gaussian Splatting
- Feed-forward reconstruction
- Recurrent refinement
- Novel view synthesis
Read analysis An academic reading of QuickSplat, covering learned Gaussian initialization, gradient-aware densification, iterative sparse-UNet optimization, ScanNet++ evidence, code correspondence, and reproducibility limits.
- 3D reconstruction
- Gaussian splatting
- Learned optimization
- Computer vision
Read analysis