Research map / 3D vision / 3D generation
Hash3D: Training-free Acceleration for 3D 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.
Hash3D accelerates diffusion-based SDS 3D generation without training by hashing and reusing intermediate U-Net features across nearby camera poses and timesteps, using adaptive grid sizes.
Problem
Score distillation sampling for 3D generation requires thousands of 2D diffusion inferences across camera views and timesteps, making optimization very slow (up to hours per object).
Approach
Observes that diffusion features from nearby camera poses and timesteps are highly similar, and stores/retrieves them via a grid-based hash table (with separate chaining and distance-weighted averaging) plus adaptive grid sizes chosen by feature cosine similarity, so SDS can skip redundant denoiser calls.
Key results
- Hash3D speeds up DreamGaussian by 4.0x (2m to 30s) and Zero-123 (NeRF) by 3.3x (20m to 7m) with comparable or slightly higher PSNR/CLIP scores.
- Across five text-to-3D methods, Hash3D yields 1.5x to 1.9x speedups with CLIP-similarity changes mostly within ±0.01.
- In a 44-participant user study, Hash3D received 52.33/100 preference against Zero-123 (NeRF) and 56.29/100 against GaussianDreamer, i.e. comparable quality.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Hash3D | DreamGaussian · GSO | time: 2m -> 30s | Helped |
| Add Hash3D | Zero-123 (NeRF) · GSO | CLIP-G/14: 0.662 -> 0.665 | Helped |
| Add Hash3D | Zero-123 (GS) · GSO | CLIP-G/14: 0.643 -> 0.632 | Hurt |
| Add Hash3D | Magic123 · GSO | time: 120m -> 74m | Helped |
| Add Hash3D | Dreamfusion · Text-to-3D | CLIP-G/14: 0.407 -> 0.411 | Helped |
| Add Hash3D | Latent-NeRF · Text-to-3D | CLIP-G/14: 0.406 -> 0.406 | No clear effect |
| Add Hash3D | SDS+GS · Text-to-3D | CLIP-G/14: 0.413 -> 0.402 | Hurt |
| Add Hash3D | GaussianDreamer · Text-to-3D | time: 15m -> 10m | Helped |
| Retrieve features without hashing (ablation) | Hash3D on Zero-123 (NeRF) · GSO | time: 7m -> 11m | Hurt |
| Hash noise prediction instead of intermediate features (ablation) | Hash3D on Zero-123 (NeRF) · GSO | CLIP-G/14: 0.665 -> 0.631 | Hurt |
| Use constant grid size (20, 20, 0.15, 20) instead of adaptive grid sizing (ablation) | Hash3D on DreamFusion · Text-to-3D | CLIP-G/14: 0.411 -> 0.345 | Hurt |
| Use temporal grid only instead of spatial and temporal grids (ablation) | Hash3D on DreamFusion · Text-to-3D | CLIP-G/14: 0.411 -> 0.280 | Hurt |
Surprising findings
- Hashing intermediate features outperforms hashing denoising predictions (noise), which significantly lowered CLIP scores.
- Using only a temporal grid without spatial grid causes large CLIP-score degradation (0.280 vs 0.411), indicating spatial redundancy is key.
Limitations
- Speedups are modest for some methods (1.5x for DreamFusion, Magic123 1.6x), and the reported MACs are theoretical assuming all retrievals succeed.
- Feature similarity degrades with large grid sizes, causing artifacts; performance depends on hyperparameters like hash probability and grid sizes.
- Evaluation of text-to-3D relies on CLIP similarity and a small user study (44 participants), which are weak proxies for 3D quality.
Experimental setup
Datasets: Google Scanned Objects (GSO) DreamFusion prompts
Benchmarks: Google Scanned Objects (GSO) novel view synthesis DreamFusion 50-prompt text-to-3D
Metrics: PSNR SSIM LPIPS CLIP similarity generation time MACs user preference
Baselines: DreamFusion Fantasia3D Latent-NeRF Magic3D GaussianDreamer DreamGaussian Zero-123 (NeRF) Zero-123 (GS) Magic123 SDS+GS
Builds on: Score Distillation Sampling DeepCache Zero-123 Stable Diffusion 3D Gaussian Splatting
Closest papers by meaning
- CAD: Photorealistic 3D Generation via Adversarial Distillation (CVPR 2024 · similarity 0.89)
- IM-3D: Iterative Multiview Diffusion and Reconstruction for High-Quality 3D Generation (ICML 2024 · similarity 0.88)
- Acc3D: Accelerating Single Image to 3D Diffusion Models via Edge Consistency Guided Score Distillation (CVPR 2025 · similarity 0.88)
- Consistent3D: Towards Consistent High-Fidelity Text-to-3D Generation with Deterministic Sampling Prior (CVPR 2024 · similarity 0.88)
- Unique3D: High-Quality and Efficient 3D Mesh Generation from a Single Image (NeurIPS 2024 · similarity 0.88)
- DreamCraft3D: Hierarchical 3D Generation with Bootstrapped Diffusion Prior (ICLR 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.