JointNet: Extending Text-to-Image Diffusion for Dense Distribution Modeling
About
We introduce JointNet, a novel neural network architecture for modeling the joint distribution of images and an additional dense modality (e.g., depth maps). JointNet is extended from a pre-trained text-to-image diffusion model, where a copy of the original network is created for the new dense modality branch and is densely connected with the RGB branch. The RGB branch is locked during network fine-tuning, which enables efficient learning of the new modality distribution while maintaining the strong generalization ability of the large-scale pre-trained diffusion model. We demonstrate the effectiveness of JointNet by using RGBD diffusion as an example and through extensive experiments, showcasing its applicability in a variety of applications, including joint RGBD generation, dense depth prediction, depth-conditioned image generation, and coherent tile-based 3D panorama generation.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Monocular Depth Estimation | NYU V2 | Delta 1 Acc86.7 | 192 | |
| Monocular Depth Estimation | ETH3D | AbsRel12.63 | 173 | |
| Monocular Depth Estimation | DIODE | AbsRel20.02 | 161 | |
| Monocular Depth Estimation | ScanNet | AbsRel14.81 | 111 | |
| Monocular Depth Estimation | KITTI | AbsRel13.74 | 92 | |
| Zero-shot affine-invariant depth estimation | ScanNet Indoor | AbsRel11.9 | 12 | |
| Zero-shot affine-invariant depth estimation | NYU Indoor v2 | AbsRel13.6 | 12 | |
| Zero-shot affine-invariant depth estimation | KITTI Outdoor | AbsRel29.9 | 9 | |
| Zero-shot affine-invariant depth estimation | ETH3D Various | AbsRel19.2 | 9 | |
| Depth-conditioned Image Generation | COCO 5k | FID25.66 | 7 |