search

justimyhxu / GRM

Large Gaussian Reconstruction Model for Efficient 3D Reconstruction and Generation
https://justimyhxu.github.io/projects/grm/
533 stars 32 forks source link

Pretrained model release #2

Open fatbao55 opened 6 months ago

fatbao55 commented 6 months ago

Thanks for the great work! Is there an estimated date for the release of the inference models used on hugging face?

tim-win commented 6 months ago

Also interested in this!

justimyhxu commented 6 months ago

Just release the inference code. Try to use it and provide us with feedback. Thank you very much!

tim-win commented 6 months ago

Transferring the models now, will be testing this out tomorrow. Going to be an exciting weekend, thank you!

fatbao55 commented 6 months ago

@justimyhxu The inference with zero123 inputs work well, thank you! However, I face OOM issues when running with SV3D inputs on a RTX3090, do you mind sharing your recommended GPU VRAM requirements?

justimyhxu commented 6 months ago

SV3D uses a latent video decoder, which results in a significant memory cost. A VRAM of 40GB works effectively with SV3D when using test.py. (Note: The Gradio demo requires more memory due to the additional models loaded for inference.)