Research map / Image generation / Fast diffusion sampling
Acc3D: Accelerating Single Image to 3D Diffusion Models via Edge Consistency Guided Score Distillation
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.
Acc3D accelerates Era3D-style single-image-to-multiview diffusion to 2-4 steps via edge-consistency-guided score distillation and disentangled adversarial regularization with the pretrained diffusion model as discriminator.
Problem
Multiview diffusion models for single-image 3D generation need many denoising steps; few-step inference degrades quality because the score function is inaccurate in the high-noise (low-SNR) endpoint region.
Approach
Apply consistency constraints only in the high-SNR "edge" region of the timestep range to refine the one-step endpoint estimate (distillation target), and combine with an adversarial loss where the pretrained diffusion model serves as discriminator, with separate discriminators for normal and texture components.
Key results
- Acc3D with 2 NFE reaches Chamfer Distance 0.0191 and IoU 0.6681 on GSO, beating Wonder3D (0.0199 / 0.6244) at 40 NFE.
- On GSO novel view synthesis at 2 steps, PSNR 22.47 vs Era3D 20.17 at 40 NFE.
- Paper claims over 20x inference speedup over the baseline multiview diffusion models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Acc3D instead of Era3D for 3D reconstruction | Era3D · GSO | CD: 0.0217 -> 0.0191 | Helped |
| Use Acc3D instead of Era3D for 3D reconstruction | Era3D · GSO | IoU: 0.5973 -> 0.6681 | Helped |
| Use Acc3D instead of Era3D for novel-view quality | Era3D · GSO | MUSIQ: 63.64 -> 67.85 | Helped |
| Use Acc3D instead of Era3D for novel-view quality | Era3D · DTC | MUSIQ: 60.13 -> 64.46 | Helped |
| Remove score distillation and adversarial learning, using consistency over [0, T] (ablation) | Acc3D · GSO | PSNR: 22.47 -> 18.11 | Hurt |
| Remove score distillation and use consistency over [0, T] (ablation) | Acc3D · GSO | PSNR: 22.47 -> 20.79 | Hurt |
| Use consistency over [0, T] instead of [0, 0.4T] (ablation) | Acc3D · GSO | PSNR: 22.47 -> 21.10 | Hurt |
| Remove consistency constraints (ablation) | Acc3D · GSO | PSNR: 22.47 -> 20.47 | Hurt |
| Remove score distillation while using consistency over [0, 0.4T] (ablation) | Acc3D · GSO | PSNR: 22.47 -> 20.32 | Hurt |
| Narrow the consistency region to [0, 0.2T] (ablation) | Acc3D · GSO | PSNR: 22.47 -> 21.22 | Hurt |
| Widen the consistency region to [0, 0.6T] (ablation) | Acc3D · GSO | PSNR: 22.47 -> 22.13 | Hurt |
| Widen the consistency region to [0, 0.8T] (ablation) | Acc3D · GSO | PSNR: 22.47 -> 21.23 | Hurt |
Surprising findings
- Ablation shows that consistency regularization over the full timestep range ([0,T]) without distillation performs worse than the restricted edge region, and adversarial training alone is unstable (mode collapse).
Limitations
- Training requires about 24 hours on four A6000 GPUs and depends on the Era3D base model.
- Edge consistency region choice (e.g., [0, 0.4T]) is a hyperparameter that significantly affects results.
Experimental setup
Datasets: Objaverse GSO DTC
Benchmarks: GSO DTC
Metrics: PSNR SSIM LPIPS MUSIQ CLIPIQA MANIQA Chamfer Distance Volume IoU
Baselines: Zero123 RealFusion Magic123 One-2-3-45 Point-E Shap-E SyncDreamer Wonder3D Era3D Stable-Fast-3D TripoSR InstantMesh
Builds on: Era3D consistency models DPM-Solver Wonder3D
Closest papers by meaning
- CAD: Photorealistic 3D Generation via Adversarial Distillation (CVPR 2024 · similarity 0.90)
- Hash3D: Training-free Acceleration for 3D Generation (CVPR 2025 · similarity 0.88)
- DreamTime: An Improved Optimization Strategy for Diffusion-Guided 3D Generation (ICLR 2024 · similarity 0.88)
- Unique3D: High-Quality and Efficient 3D Mesh Generation from a Single Image (NeurIPS 2024 · similarity 0.88)
- Sparse3D: Distilling Multiview-Consistent Diffusion for Object Reconstruction from Sparse Views (AAAI 2024 · similarity 0.87)
- ACT-Diffusion: Efficient Adversarial Consistency Training for One-step Diffusion Models (CVPR 2024 · similarity 0.87)
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.