Tech News

NVIDIA announced an AI capable of creating three-dimensional objects using photographs

Artificial intelligence GET3D. (photo: YouTube/NVIDIA)

After the “boom” of the images produced by the artificial intelligencethe next step of this technology is appearing little by little to the delight of all users.

NVIDIA announced GET3D, a new AI model that can create 3D objects that can be used in video game, films or in the metaverse. Using the GPU and 2D images taken from multiple angles, the model can create complex shapes with high-quality textures.

According to NVIDIA’s blog post, engineers trained the AI ​​model on 1 million images. Using multiple A100 Tensor Core GPUs, the team reached their goal in two days. According to the company, GET3D can produce 20 objects per second with a single GPU, which would facilitate the work of artists and content producers.

“The real world is full of variety: the streets are full of unique buildings, with different vehicles buzzing by and diverse crowds passing by. Manually modeling a 3D virtual world that reflects this is time consuming, making it difficult to complete a detailed digital environment.

Until now, GET3D has been able to create various objects, from cars, chairs, animals or motorcycles to people or buildings. The template can be combined with other NVIDIA AI tools to create an object with a specific style.

Engineers mention that it can create highly detailed 3D networks with complex topology and realistic textures that can be exported to rendering engines. games either Applications rendering.

3D objects are editable for use in movies or video games

One of the main points of GET3D is that the results can be modified, which was difficult in previous experiments.

At first glance, the result is similar to what can be obtained when using the photogrammetry technique with a detailed network and structure model. The only drawback is that models of this class require some subsequent “cleanup” work to optimize the number of polygons.

“Previous work on 3D generative modeling lacks geometric details, is limited in the mesh topology they can produce, typically doesn’t support textures or uses neural renderers in the synthesis process, making its use in common 3D software unsuitable. trivial”, can be read on the official NVIDIA blog.

As with other deep learning models, the larger the training data, the better the results. GET3D uses a synthetic dataset, although the researchers plan to train the next version on real data.

NVIDIA GET3D: AI model to populate virtual worlds with 3D objects and characters

The next version of GET3D is already in the plans

Creating 3D objects using artificial intelligence is interesting and NVIDIA has taken an important step. However, 3D models for video games, movies, or other entertainment applications are much more complex than a Mindjourney or DALL-E experiment.

These objects must meet additional properties such as polygon number or topology when used in games or scenes that contain animations.

(photo: Kaedim)
(photo: Kaedim)

A few weeks ago, a similar show was trending on Twitter. The service Kaedimwhich uses artificial intelligence to create 3D models from 2D images, It has caught the attention of several artists. After publishing impressive results that were too good to be true, some accused the company of lying and using 3D professionals.

Finally, Kaedim appeared and said that although he uses an AI algorithm to create the object, a human changes the end result. A “quality control engineer” checks the model and adjusts it to the quality standard required by the customer before shipment.

Although it optimizes production time compared to manual work, the algorithm still cannot provide reliable results.

Related Articles

Back to top button