Nvidia is doing pretty well. After publishing the Blackwell superchip designed for training more powerful AI models such as GPT, Claude, and Gemini, he published his own Text-to-3D AI tools (great for consumer options) (See our guide on graphics cards). .
The graphics card giant wrapped up GTC week by showcasing LATTE3D, a text-to-3D generative AI model it described as a “virtual 3D printer.” It can convert text prompts into his 3D representations of objects and animals in less than a second.
Nvidia says the 3D shapes generated by LATTE3D can be “easily delivered in virtual environments such as video game development, advertising campaigns, design projects, or robotics virtual training grounds.” We've seen tools that convert text to 3D before, but according to online praise, some people aren't too impressed with the quality of LATTE3D's results. However, the new model represents a significant improvement, especially in terms of speed.
According to Nvidia, running inference on a single GPU, such as the NVIDIA RTX A6000 used in the research demo, generates 3D shapes almost instantly. This means that a creator starting a design from scratch or exploring a 3D asset library can use LATTE3D to generate detailed objects as soon as an idea comes to mind.
The model generates several 3D shape options based on each text prompt. High-quality optimization of desired objects and export to graphics software applications and platforms such as NVIDIA Omniverse enables Universal Scene description (OpenUSD)-based 3D workflows and applications.
“A year ago, it took an hour for an AI model to produce a 3D visual of this quality, whereas with today's state-of-the-art technology it takes about 10 to 12 seconds,” said Sanja Fidler, vice president of AI research. I am. Results are generated orders of magnitude faster, bringing near real-time text-to-3D generation to creators across all industries. ”
LATTE3D was developed by Nvidia's Toronto-based AI Labs team and trained using text prompts generated using ChatGPT, asking users to come up with different phrases to describe a particular 3D object. Improved the ability of the model to process. The researchers trained his LATTE3D on two specific datasets: animals and everyday objects, but the same architecture can also be used to train AI on other data types. This is a research project only and is not available for general use.
AI creator Bilawal Sidhu writes: X: “This breakthrough is huge. DreamFusion around 2022 was slow and low-quality, but it started this generative 3D revolution. Initiatives like ATT3D (Amortized Text to 3D Object Composition) We sacrificed quality for speed. Now with LATTE3D, quality and processes take less than a second! That means you can use text or images in 3D to quickly iterate through your 3D world. This means that you can set it. ”
Along with video, 3D is the next frontier for AI image generation. Also this week, Adobe announced that it will be integrating its first Firefly AI-powered tool into Substance 3D.