Research map / 3D vision / 3D generation
Turbo3D: Ultra-fast Text-to-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.
Turbo3D distills a multi-step multi-view text-to-image diffusion model into a 4-step, 4-view student using dual-teacher DMD (multi-view plus single-view teachers), then reconstructs 3D Gaussians with a latent-space GS-LRM for sub-second text-to-3D generation.
Problem
Text-to-3D generation via multi-view diffusion and reconstruction is slow due to iterative denoising, and naive distillation of multi-view models causes mode collapse toward synthetic Objaverse-style appearance.
Approach
A dual-teacher distribution matching distillation, combining a multi-view DMD loss (for consistency) with a single-view DMD loss (for photorealism), plus a GS-LRM that consumes multi-view latents directly instead of decoded pixels.
Key results
- Turbo3D achieves CLIP score 27.61 and VQA score 0.76 in 0.35s on A100, versus Instant3D 26.23/0.65 at 15.02s and LGM 24.73/0.58 at 6.56s.
- In a user study, Turbo3D won 89.8% against LGM, 74.9% against Instant3D, and 50.6% against its MV teacher.
- Dual-teacher distillation raises CLIP score from 26.60 (MV teacher only) to 27.61, close to the multi-step MV model's 28.04; latent GS-LRM reduces inference time from 0.45s to 0.35s with no quality loss.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| distill with both the MV and single-view teachers rather than only the MV teacher (ablation) | Turbo3D few-step generator · DreamFusion prompts | CLIP Score: 26.60 -> 27.61 | Helped |
| distill with both the MV and single-view teachers rather than only the MV teacher (ablation) | Turbo3D few-step generator · DreamFusion prompts | VQA Score: 0.69 -> 0.76 | Helped |
| distill the MV teacher into a few-step generator (ablation) | Turbo3D · DreamFusion prompts | CLIP Score: 28.04 -> 26.60 | Hurt |
| distill the MV teacher into a few-step generator (ablation) | Turbo3D · DreamFusion prompts | VQA Score: 0.77 -> 0.69 | Hurt |
| reconstruct directly from multi-view latents instead of using pixel-space GS-LRM (ablation) | Turbo3D · DreamFusion prompts | CLIP Score: 27.62 -> 27.61 | No clear effect |
| reconstruct directly from multi-view latents instead of using pixel-space GS-LRM (ablation) | Turbo3D · DreamFusion prompts | Inference time: 0.45s -> 0.35s | Helped |
| use Turbo3D instead of Instant3D | text-to-3D generation · DreamFusion prompts | CLIP Score: 26.23 -> 27.61 | Helped |
| use Turbo3D instead of Instant3D | text-to-3D generation · DreamFusion prompts | Inference time: 15.02s -> 0.35s | Helped |
Surprising findings
- Naively distilling the multi-view teacher causes compound mode collapse toward cartoonish Objaverse-style appearance, which a single-view teacher corrects.
Limitations
- Multi-view teacher and generator fine-tuned on synthetic Objaverse data, which biases appearance toward synthetic style.
- Distilled student falls below the multi-step MV model in CLIP and VQA scores (27.61 vs 28.04).
Experimental setup
Datasets: Objaverse Cap3D captions
Benchmarks: DreamFusion 400 prompts
Metrics: CLIP score VQA score inference time user study win rate
Baselines: Instant3D LGM TripoSR SV3D MV teacher (multi-step)
Builds on: Distribution Matching Distillation DMD2 GS-LRM Instant3D Objaverse
Closest papers by meaning
- DiffSplat: Repurposing Image Diffusion Models for Scalable Gaussian Splat Generation (ICLR 2025 · similarity 0.90)
- Instant3D: Fast Text-to-3D with Sparse-view Generation and Large Reconstruction Model (ICLR 2024 · similarity 0.90)
- Text-to-3D using Gaussian Splatting (CVPR 2024 · similarity 0.89)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.89)
- IM-3D: Iterative Multiview Diffusion and Reconstruction for High-Quality 3D Generation (ICML 2024 · similarity 0.89)
- DirectTriGS: Triplane-based Gaussian Splatting Field Representation for 3D Generation (CVPR 2025 · 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.