How to create higher resolution(1024) result

[hayoung-jeremy]

I've successfully installed and run the code, and it's amazing!
Thank you guys for the great work :)

btw, I've seen on the paper says :

Notably, we maintain the fast speed to generate 3D objects within 5 seconds while boosting the training resolution to 512, thereby achieving high-resolution 3D content generation.

Is there an way to produce higher resolution such as 1024?

I've tried to set the option as follows, but it produced cracked result :

python infer.py big --input_size 1024 --output_size 1024 --splat_size 512 --resume pretrained/model.safetensors --workspace result --test_path /home/ubuntu/altava-2/input-images/test.png
python convert.py big --test_path result/test.ply --input_size 1024 --output_size 1024 --splat_size 512 --force_cuda_rast

the result glb :

original glb :

[ashawkey]

@hayoung-jeremy Hi, since the model is trained on 512x512, it may fail to infer with a different resolution. You'll need to train a new model using 1024x1024 output resolution, but this may lead to too many 3D Gaussians and consume lots of GPU memory.

[hayoung-jeremy]

Thank you for your kind response!
If I'm looking for higher resolution outputs, similar to 1024, what technology would you recommend using?
Also, if it's okay to use a lot of 3D Gaussian and GPU memory, would it be alright to proceed with training?
Could you also inform me about the estimated time required for generation?