Acc3D: Accelerating Single Image to 3D Diffusion Models </br> via Edge Consistency Guided Score Distillation

Our method completed the Image-to-3D generation process within 1.5 seconds while achieving superior generation quality.

Videos are played at the original speed and recorded on an NVIDIA A6000.

We present Acc3D to tackle the challenge of accelerating the diffusion process for generating 3D models from single images. To derive accurate reconstruction through few-step inference, we emphasize the critical issue as the modeling of the score function at the endpoints (states of the random noise). To tackle such an issue, we propose edge consistency, i.e., consistent predictions across the low signal-to-noise ratio region, to enhance a pre-trained diffusion model, enabling a distillation-based refinement of the endpoint score function. Building on those distilled diffusion models, we introduce an adversarial augmentation strategy to further enrich generation detail. The two modules complement each other, mutually reinforcing to elevate generative performance. Extensive experiments show that our Acc3D not only achieves over a 20x increase in computational efficiency but also yields notable quality improvements, compared with state-of-the-art methods.

Overview of our Acc3D. The training pipeline unfolds in two core components: edge consistency-guided distillation and adversarial training. Each component bolsters the other's advantages—the distillation procedure stabilizes adversarial training, mitigating the risk of mode collapse, while adversarial learning can enhance perceptual richness. Collectively, these elements craft a balanced, refined model that excels in both stability and detail.

We visualized the comparison with Era3D's results, and our accelerated model outperforms Era3D significantly.

our method provides a generic acceleration framework for diffusion models. It also achieves good performance with fewer steps in image generation tasks.