There are a lot of eyes on NVIDIA's Neuralangelo that just this past week was named one of Time Magazine's Top Inventions of 2023. Instant-Angelo comes from CUHK-Shenzhen MPhil Student, Chongjie Ye.
It wasn't the greatest time for Instant-Angelo to be released, but the product will speak for itself.
Just like how Instant-NeRF sped up workflows by a magnitude, Instant-Angelo is primed to do the same. The resulting method cuts down the processing time from roughly 40 hours on a A100 to 20 minutes. While Neuralangelo will also take a large amount of VRAM to run, Instant Angelo is able to run with a minimum of 10gb of VRAM, expanding its capability to consumer graphics cards.
There are two reconstruction modes that are able to be utilized: Smooth and Detailed. Smooth takes roughly 20 minutes to run, whereas Detailed is a bit more involved, coming in at roughly an hour. Like the name implies, Smooth is better for objects or scenes that doesn't necessarily have a ton of intricate details.
As a note, Smooth does have the potential to oversmooth certain areas of flat areas or slight changes of elevations.
For Detail mode, they suggest using it in the following situations:
Image data captured under varying conditions over time or with inconsistent exposure levels.
High resolution image sources of 2K or 4K dimensions.
Your images' resolution are 2K or 4K
Reconstructing objects or scenes comprised of glossy, reflective materials.
Subjects containing large textureless or untextured surface regions.
While it's still not instantaneous, it is an order of magnitude faster and great that it's able to be run with consumer grade GPUs.
Currently, it is only been tested on a Linux machine, though it's not impossible that it won't work on Windows. To give it a try, check out their Github page!
