Improving NeRF Models with Super-Resolution for Faster Inference
Super-resolution has been proposed to enhance the outputs of neural radiance fields (NeRF) and generate high-quality images at faster speeds. Existing NeRF+SR methods have added complexity and training overhead, making them less efficient. In this paper, a simple NeRF+SR pipeline is proposed, along with a lightweight augmentation technique, random patch sampling, for training.
Efficient NeRF+SR Pipeline
The proposed NeRF+SR pipeline combines existing modules to mitigate the computing overhead for super-resolution. It can be trained up to 23× faster, making it feasible to run on consumer devices such as the Apple Macbook.
Improved Inference Speeds
Experiments show that the pipeline can upscale NeRF outputs by 2-4× while maintaining high quality and increasing inference speeds by up to 18× on an NVIDIA V100 GPU and 12.8× on an M1 Pro chip.
These results suggest that super-resolution can be a simple and effective technique for improving the efficiency of NeRF models for consumer devices.