GTC: Shutterstock Teams with NVIDIA to Build AI Models for Generative 3D

There have been many interesting Machine Learning based announcements at GTC. For example, Shutterstock a leading creative platform and stock library, and NVIDIA today announced that they are partnering to train a machine learning platform to create custom 3D models with generative 3D tools from simple text prompts. 3D models will be trained with Shutterstock assets using NVIDIA Picasso generative AI cloud services to convert text into high-fidelity new 3D content, reducing creation time from hours to minutes.

The new text-to-3D models will be introduced in the coming months on Shutterstock.com, as part of their Creative Flow toolkit. The text-to-3D features will also be offered on Turbosquid.com and are planned to be introduced on the NVIDIA Omniverse platform for use in building and operating 3D industrial metaverse applications.

Building high-quality, detailed 3D models from scratch is often a challenging and time-consuming task. This is true in VFX as it is in industrial applications such as use in a digital twin simulation. The 3D often requires precision and the complexity of the job can take days or even longer depending on the specifics of the model. By creating custom models with the NVIDIA Picasso generative AI service, Shutterstock will help 3D artists to ideate and explore new forms in minutes. NVIDIA Picasso is a cloud service for building and deploying generative AI-based image, video, and 3D applications. It’s optimized to deliver the highest training and inference performance for generative AI visuals via the NVIDIA DGX super computing cloud.

Artists, TDs, and Animators can create object shapes, assist with object unwrapping, generate textures and materials at the high-end, and for non-power 3D users, the system will allow for new 3D models to be easily used in a wide variety of applications and platforms.

Initially, the system will produce just one mesh per object and with only a simple UV color texture. In the future, the system will be expanded to include more complex sub-divisions of the model into logical component pieces and more advanced texturing options. While Shutterstock’s Turbosquid is an invaluable digital backlot of assets, clearly the current library favors creating commonly used 3D objects. The new text-to-3D will allow for complex bespoke new models to be made. Building such unusual or specialist models would be expensive and time-consuming. With modern agile project approaches, it will be possible to iterate designs to a client quickly before taking the base model and then refining it further to a polished detailed textured asset.

The transformative capabilities of generative AI make it possible for software makers and enterprises to build tools that use simple text prompts to create 3D assets for digital twins, simulation, and design, saving artists enormous amounts of time and effort,” said NVIDIA Vice President of Developer Programs Greg Estes. “Training a custom Shutterstock model with the NVIDIA Picasso generative AI cloud services will give developers a tool that can automate much of the tedious work for artists, freeing them to spend more time exploring new concepts and refining their ideas.”

To be ethical and fair to artists the combined text-to-3D plan has been built to not ‘steal’ work from artists nor allow people to sell assets made generatively through the normal Turbosquid marketplace. All artists can opt out of their work being used for any training purposes and both companies are keen to make sure this helps freelance 3D artists and treats them fairly. As part of its responsible AI focus and in correlation with sales of the customized 3D models on Shutterstock’s platform, Shutterstock will compensate artists through its Contributor Fund for the role their IP plays in training the generative technology.