When will the Hugging Face demo experience be provided? How much graphics memory is needed to run, how long does it take to generate, and how does the effect compare to Unique3D

[libai-lab]

When will the Hugging Face demo experience be provided? How much graphics memory is needed to run, how long does it take to generate, and how does the effect compare to Unique3D

[ch3cook-fdu]

Thanks for your interest in our work! We will make that happen in the near future.

The inference time and memory use depend heavily on the length of sequences and the model size. Because of the O(N^2) time and memory complexity of the self-attention mechanism, the memory cost and time consumption are currently the bottleneck of our work. Please see our paper for further discussions.

Difference with Unique3D. As far as I am concerned, Unique3D adopts a multi-view reconstruction approach for 3D content creation. Our MeshXL differs from Unique3D in that we directly builds a family of generative 3D foundation models directly with large scale 3D data by treating coordinates as languages.