Generation time

[guochengqian]

Generation time

[ecumene]

Inference+optimization takes approximately 1 hour on an A100 from what I've heard (source https://twitter.com/Eng_Hemdi)

[guochengqian]

Around (less than) 1 hour on an 16G v100 with preprocessed foreground segmentation, depth map, and textural inversion.
We will provide the preprocessed files for all examples we tested on.

For a new image, the foreground segmentation and depth map extraction takes seconds, but the textural inversion takes 2 hours on a 32G v100. Textural inversion is the bottleneck, you might need faster textural inversion if there exists any. Or simply using a pure text prompt with reduced texture consistency.

[reynoldscem]

Great, thank you!

[yuedajiong]

the GPU requirement is too high, and the infer time is too long.
I want to modified it.