Adobe demonstrates new AI large model to achieve 2D to 3D conversion in just 5 seconds

What needs to be rewritten is: Virtual Reality Gyro

Recently, Adobe Research collaborated with researchers from the Australian National University to develop a new AI large model called LRM. This model converts 2D images into high-quality 3D models in just 5 seconds

Unlike past methods of training on small datasets in a category-specific manner, LRM adopts a highly scalable Transformer-based neural network architecture with more than 500 million parameters. The model is trained in an end-to-end manner on approximately 1 million 3D objects from Objaverse and MVImgNet datasets and predicts Neural Radiation Fields (NeRF) directly from the input images

Source: yiconghong.me/LRM/

The paper "LRM: LARGE RECONSTRUCTION MODEL FOR SINGLE IMAGE TO 3D" states that "the combination of large-capacity models and large-scale training data makes our model highly versatile and performs well according to a variety of test inputs (including real-world Capture and generate images of the model) to produce high-quality 3D reconstructions."

Experiments have shown that LRM can reconstruct high-fidelity 3D models based on real-world images and images created by AI-generated models such as DALL-E and Stable Diffusio. The system can generate detailed geometric shapes and preserve complex textures such as wood grain. However, LRM still has certain limitations in texture generation in occluded areas.

In terms of application, LRM has broad prospects. It can be applied to various scenarios such as industrial design, games and entertainment. In addition, in the field of AR/VR, LRM can also improve user experience by generating detailed 3D environments in real time

Without changing the original meaning, the content that needs to be rewritten into Chinese is: Source: venturebeat

The above is the detailed content of Adobe demonstrates new AI large model to achieve 2D to 3D conversion in just 5 seconds. For more information, please follow other related articles on the PHP Chinese website!