Solved

[muses0229]

No description provided.

[imlixinyang]

We have tested some cases on a single RTX 3090. The memory cost is very close to the maximum so we gave a simple solution. You can try running the refining separately to avoid OOM. For example:

1. generate camera and 3DGS without refining:

python inference.py --export_all --text '{text}' --num_refine_steps 0 --num_samples 4

2. see the results in exps/tmp/videos and choose a sample (filename) for separate refining:

python refine.py --ply 'exps/tmp/ply/{filename}.ply'  --camera 'exps/tmp/camera/{filename}.npy' --export_all --text '{text}' --num_refine_steps 1000

This has been tested on a single T4 GPU (16 GB). Let me know if it works!

[githubnameoo]

We have tested some cases on a single RTX 3090. The memory cost is very close to the maximum so we gave a simple solution. You can try running the refining separately to avoid OOM. For example:

1. generate camera and 3DGS without refining:

python inference.py --export_all --text '{text}' --num_refine_steps 0 --num_samples 4

2. see the results in exps/tmp/videos and choose a sample (filename) for separate refining:

python refine.py --ply 'exps/tmp/ply/{filename}.ply'  --camera 'exps/tmp/camera/{filename}.npy' --export_all --text '{text}' --num_refine_steps 1000

This has been tested on a single T4 GPU (16 GB). Let me know if it works!

How much gpu memory is actually needed to test?

[imlixinyang]

Not a sure number. Since the number of Gaussian points varies for different scenes during refining, the GPU cost also varies.
Typically, 28GB is basically enough for running the refining jointly with generation, and 16GB is enough for running the refining separately.
@githubnameoo