search

lutao2021 / BrightDreamer

BrightDreamer: Generic 3D Gaussian Generative Framework for Fast Text-to-3D Synthesis
https://vlislab22.github.io/BrightDreamer/
MIT License
83 stars 2 forks source link

About out-of-domain text-to-3D generation. #4

Closed Colezwhy closed 6 months ago

Colezwhy commented 6 months ago

Hi, thank you for your great work, which is very inspiring to me. But I was wondering if we use prompts that are out of the training set domain(i.e. vehicle, animal, daily life), what will Bright Dreamer produce? Will it generalize to other domains a little bit?

lutao2021 commented 6 months ago

Hi, thanks for your focus.

Though it can generate 3D content for unseen text, it still requires the text not to deviate far from the training set. If you use the prompt which is out of training set domain, it cannot generate the semantic-consistent content. Actually, we also attempt to train BrightDreamer using a more abundant prompts set.

I think if you can prepare high-quality and high-richness prompt sets, it can generalize to other domains.

Colezwhy commented 6 months ago

Hi, thanks for your focus.

Though it can generate 3D content for unseen text, it still requires the text not to deviate far from the training set. If you use the prompt which is out of training set domain, it cannot generate the semantic-consistent content. Actually, we also attempt to train BrightDreamer using a more abundant prompts set.

I think if you can prepare high-quality and high-richness prompt sets, it can generalize to other domains.

Thank you for your reply! So when confronting more generalized scenarios, it still needs a bigger dataset to obtain its generalizability. BTW, do you plan to release a model that is trained on a bigger dataset? It might help in the field of reducing time cost for 3D gen. Good job! Congrads!

lutao2021 commented 6 months ago

Sorry. Due to other work I am currently doing, I am unable to provide an accurate date. But we will attempt it in the next few months.

Colezwhy commented 6 months ago

Sorry. Due to other work I am currently doing, I am unable to provide an accurate date. But we will attempt it in the next few months.

Thank you, I will close this.