NVIDIA announces AI ``Magic3D'' that generates high-resolution 3D models from text, it is also possible to fine-tune text and imitate styles



NVIDIA, a major semiconductor manufacturer and also focusing on AI development, has announced AI ` ` Magic3D '' that generates 3D models based on input text. The 3D model generated by Magic3D boasts eight times the resolution of '

DreamFusion ' announced by Google Research, and takes about half the time to generate.

[2211.10440] Magic3D: High-Resolution Text-to-3D Content Creation
https://doi.org/10.48550/arXiv.2211.10440

Magic3D: High-Resolution Text-to-3D Content Creation
https://deepimagination.cc/Magic3D/

3D for everyone? Nvidia's Magic3D can generate 3D models from text | Ars Technica
https://arstechnica.com/information-technology/2022/11/nvidias-magic3d-creates-3d-models-from-written-descriptions-thanks-to-ai/

Nvidia's Magic3D turns text into high-resolution 3D objects
https://the-decoder.com/nvidias-magic3d-turns-text-into-high-resolution-3d-objects/

Magic3D is an AI that generates high-resolution 3D models based on input text (prompts), similar to various image generation AIs. When you enter the prompt 'A blue poison-dart frog sitting on a water lily.', a 3D model of a poisonous blue frog sitting on a leaf is generated. will be



A colorless 3D mesh was also shown.



It is like this when it is 'A silver platter piled high with fruits.'



In addition, it seems that you can generate 3D models based on various texts.



Each 3D mesh looks like this.



Also, by changing part of the prompt that generated the 3D model in the first place, it is possible to add various variations to the same composition. The following 3D models are, from left to right, ``a baby rabbit sitting on a heaping pancake'', ``a metal rabbit sitting on a heaping broccoli'', and a ``a sphinx sitting on a heaping chocolate cookie''. .



By entering a reference image, it is also possible to imitate the style of the image and generate a 3D model.



Magic3D employs a two-step process to generate 3D models. First, based on the input text, a 2D image is generated using '

eDifi ', a highly accurate image generation AI.



Next, NVIDIA's

Instant-NGP is used to generate a low-resolution 3D model from the 2D image.



The next step is to extract a high-resolution 3D model from a low-resolution 3D model using

DMTet AI , which synthesizes a high-resolution 3D model from a coarse 3D mesh. The DMTet used by Magic3D is optimized for this purpose.



NVIDIA explains that Magic3D generates high-resolution 3D models with this method.



NVIDIA is also comparing DreamFusion and Magic3D, a 3D model generation AI announced by Google Research. Below is a 3D model generated by DreamFusion (left) and a 3D model generated by Magic3D (right) based on the text 'A plate piled high with chocolate chip cookies.' Something.



A comparison of the 3D models generated by DreamFusion and Magic3D with the text 'Michelangelo style statue of an astronaut.' on the left and 'A ceramic lion.' on the right. . Certainly Magic3D can generate higher resolution 3D models. According to NVIDIA, Magic3D can generate 3D models with eight times higher resolution than DreamFusion, and DreamFusion takes an average of 1 hour and 30 minutes to generate, while Magic3D averages about 40 minutes.



NVIDIA's research team hopes that anyone will be able to create 3D models without special training. As Magic3D becomes more sophisticated, it could speed up the development of games and VR content, and eventually find its way into film and television special effects. In the paper, the research team said, ``We hope that Magic3D will be able to popularize 3D model synthesis and unleash everyone's creativity in 3D content production.''

in Software,   Science, Posted by log1h_ik