Research map / Image generation / Fast diffusion sampling
DreamPropeller: Supercharge Text-to-3D Generation with Parallel Sampling
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.
DreamPropeller generalizes Picard iterations to parallelize score-distillation text-to-3D pipelines (SDS/VSD over NeRF, DMTet, SDF, 3D Gaussian Splatting), achieving about 4.2-4.7x wall-clock speedup at similar quality.
Problem
Score-distillation text-to-3D generation (SDS/VSD) is very slow because gradient-based optimization of 3D representations requires many sequential iterations, and standard Picard iteration cannot handle Adam momentum or changing parameter dimensions.
Approach
Generalize Picard iterations to non-ODE sequential computations (Adam momentum updates and dimension-changing 3D Gaussian splitting/pruning) via a pseudo-inverse formulation, parallelize the expensive SDS/VSD gradient computations across GPUs in a sliding window with an adaptive EMA convergence threshold.
Key results
- Achieves 4.22x (DreamFusion), 4.17x (Magic3D), 4.18x (TextMesh), 4.67x (DreamGaussian), and 4.69x (ProlificDreamer) wall-clock speedup on 8 A100 GPUs.
- ProlificDreamer R-Precision rises from 94.11 to 96.12 with DreamPropeller, while FID_CLIP stays about the same (49.66 vs 49.65).
- Speedup peaks at window size 7 (one less than 8 GPUs) and grows with batch size, benefiting VSD most.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Wrap DreamFusion with DreamPropeller | DreamFusion · DreamFusion gallery (30 prompts) | Runtime: 8080 -> 1910 s; 4.22x | Helped |
| Wrap Magic3D with DreamPropeller | Magic3D · DreamFusion gallery (30 prompts) | Runtime: 13470 -> 3230 s; 4.17x | Helped |
| Wrap TextMesh with DreamPropeller | TextMesh · DreamFusion gallery (30 prompts) | Runtime: 7690 -> 1830 s; 4.18x | Helped |
| Wrap DreamGaussian with DreamPropeller | DreamGaussian · DreamFusion gallery (30 prompts) | Runtime: 700 -> 150 s; 4.67x | Helped |
| Wrap ProlificDreamer with DreamPropeller | ProlificDreamer · DreamFusion gallery (30 prompts) | Runtime: 7390 -> 3710 s; 4.69x | Helped |
| Wrap DreamFusion with DreamPropeller | DreamFusion · DreamFusion gallery (30 prompts) | R-Precision: 82.70 -> 79.57 | Hurt |
| Wrap Magic3D with DreamPropeller | Magic3D · DreamFusion gallery (30 prompts) | R-Precision: 88.45 -> 87.18 | Hurt |
| Wrap TextMesh with DreamPropeller | TextMesh · DreamFusion gallery (30 prompts) | R-Precision: 88.06 -> 84.43 | Hurt |
| Wrap DreamGaussian with DreamPropeller | DreamGaussian · DreamFusion gallery (30 prompts) | R-Precision: 83.28 -> 83.85 | Helped |
| Wrap ProlificDreamer with DreamPropeller | ProlificDreamer · DreamFusion gallery (30 prompts) | R-Precision: 94.11 -> 96.12 | Helped |
| Use a window size of 7 rather than smaller or larger windows (ablation) | DreamPropeller · DreamFusion gallery (30 prompts) | Speedup | Mixed |
| Use adaptive thresholding with γ < 1 instead of γ = 1 (ablation) | DreamPropeller · DreamFusion gallery (30 prompts) | Speedup | Helped |
Surprising findings
- DreamPropeller slightly improved R-Precision for ProlificDreamer (94.11 to 96.12) despite being an acceleration method.
- DreamFusion and TextMesh show a noticeable drop in R-Precision (82.70 to 79.57; 88.06 to 84.43) after acceleration.
Limitations
- Results do not exactly match baselines because fixed-point error is not fully converged and Adam leads to different local optima.
- Requires multiple GPUs (one pretrained diffusion model per GPU) to achieve the speedup.
- Speedup decreases for window sizes beyond the number of GPUs.
Experimental setup
Datasets: DreamFusion gallery prompts (30) ImageNet 2012 validation set (for FID reference)
Benchmarks: DreamFusion gallery prompts (30 prompts)
Metrics: CLIP R-Precision CLIP FID wall-clock runtime speedup
Baselines: DreamFusion Magic3D TextMesh DreamGaussian ProlificDreamer
Builds on: Picard iterations Shih et al. parallel sampling of diffusion models Score Distillation Sampling Variational Score Distillation 3D Gaussian Splatting
Closest papers by meaning
- DreamFlow: High-quality text-to-3D generation by Approximating Probability Flow (ICLR 2024 · similarity 0.90)
- Walking the Schrödinger Bridge: A Direct Trajectory for Text-to-3D Generation (NeurIPS 2025 · similarity 0.89)
- Rethinking Score Distilling Sampling for 3D Editing and Generation (ICML 2025 · similarity 0.89)
- LucidDreamer: Towards High-Fidelity Text-to-3D Generation via Interval Score Matching (CVPR 2024 · similarity 0.89)
- Consistent3D: Towards Consistent High-Fidelity Text-to-3D Generation with Deterministic Sampling Prior (CVPR 2024 · similarity 0.88)
- Consistent Flow Distillation for Text-to-3D Generation (ICLR 2025 · 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.