The Next Dimension of Generative AI
Portrait:
The old adage “a picture is worth a thousand words” shows why there’s currently such an explosion of interest in and excitement around AI. It started as enthusiasm for AI text generation (i.e., text-to-text; e.g., ad copy generation) and its potential, and has escalated to fervor as advances in image generation lead to a flurry of compelling showcases and proofs-of-concept.
We’re intrigued by these eye-catching demonstrations of how models can reduce barriers to creating everything from digital portraits to interior design ideas. While experimenting with text-to-image interfaces, we were also struck by the difficulty of generating content requiring a high degree of precision. It took us a few prompts in Stable Diffusion before producing the beautiful Pensieve image we used in Generative AI: Context Is All You Need.
These exciting interfaces are primarily generating 2D content, which begs the question, If a picture is worth a thousand words, then how many is a video worth? We believe 3D content is not only valuable, but also a great place for startups to build. 3D content generation is much more difficult and time-consuming to create. Additionally, there are fewer people that have the necessary skills to produce it.
Landscape:
We believe AI can contribute to huge improvements in scaling 3D content generation. We have sought over the past few years to identify companies working on problems that require fast and precise generation of 3D assets and models. A few use cases that we come across frequently include 3D for digital commerce, 3D for simulation, and 3D for manufacturing.
For example, we invested in Hexa in 2020. Hexa uses generative models to produce high-fidelity 3D counterparts that can be used for digital commerce. Large digital retailers rely on Hexa’s technology to give customers the option to view potential purchases in 3D, and game developers use Hexa to populate virtual worlds with precise replicas of real-world objects. With each generated asset, Hexa collects valuable information such as object shape, surface texture, and color that the company uses to improve model performance, making it cheaper and faster to generate content over time.
Like Hexa, Point72 Ventures’ portfolio company Blackshark uses models to generate 3D assets from 2D data. While Hexa is focused on objects, Blackshark uses generative models to reconstruct assets at planetary scale. Traditional 3D applications required hand modeling or photogrammetry solutions to create digital twins, but Blackshark is able to transform 2D imagery into synthetic 3D scenes. In less than 72 hours of processing time, Blackshark reconstructs the entire planet including its billions of buildings and billions of vegetation acres in annotated 3D [VentureBeat, 2022]. By creating accurate and multi-faceted 3D worlds, the company can enable immersive user experiences like Microsoft’s Flight Simulator.
Finally, we anticipate companies will build generative models that can simulate complex, time-consuming, and expensive processes. For example, Atomic Industries is building models to help manufacturers generate designs for tool and die making – a critical process in producing plastic objects. Traditionally this process requires skilled tool engineers to iterate on designs for months until they arrive at one that meets requirements. Atomic’s technology would rapidly generate designs and evaluate each iteration using simulation and models, which could significantly reduce design time and engineer involvement. In summary, Atomic believes that its generative design technology will help reduce the complexity of tool design and alleviate a major bottleneck in manufacturing.
We’re excited about these businesses because they demonstrate how generative models could solve pressing enterprise problems in specific verticals. As these model-driven businesses expand their engagements with early adopters, they collect valuable data that can be used to further refine their generative capabilities. Over time, we expect their models to improve to the point where they can efficiently serve the long tail of customers.
Action Shot:
As generative models continue to improve their ability to generate image and video content over the coming months, we anticipate startups will tackle novel use cases that were not previously possible. For example, we’re eager to speak with individuals considering generative AI solutions for film and television, or founders building tools to enable low latency, immersive 3D environments in the gaming space.