Unique3D is a state-of-the-art model for 3D mesh generation from single images, noted for its efficiency and high fidelity. Unique3D is code and weights are available open source. This approach is producing detailed and consistent 3D models in under 30 seconds, outperforming existing methods. This speed and quality make it a powerful tool for applications requiring rapid and detailed 3D model generation.
Unique3D Model
The Unique3D pipeline begins with a single wild image as input. First, four orthographic multi-view images are generated using a multi-view diffusion model. These images are then progressively upscaled through a multi-level process. Next, a normal diffusion model generates corresponding normal maps, which are similarly upscaled. Finally, the high-resolution color images and normal maps are used to reconstruct high-quality 3D meshes using the ISOMER algorithm, ensuring instant and consistent results.
The image generation process begins with the Stable Diffusion Image Variations Model, progresses through an upscaled version fine-tuned from ControlNet-Tile, and culminates with the pre-trained Real-ESRGAN model. The initial stage of normal map prediction also leverages the Stable Diffusion Image Variations model.
Dataset
The dataset employs a subset of the Objaverse dataset, refined to approximately 50k objects by filtering out scenes with multiple objects, low-resolution imagery, and unidirectional faces. To handle surfaces without thickness, eight orthographic projections are rendered per object, identifying 13k instances of invalid data through epipolar line analysis. The dataset is augmented with random environment maps and lighting, ensuring robustness. All images are rendered at 2048 × 2048 pixels for high-quality generation.
Results
The ISOMER algorithm’s integration of color and geometric priors is key to achieving high-quality textures and geometric accuracy.
Performance Comparison
Unique3D outperforms other methods in various metrics. Here is a comparison table:
