Result is not sharp on lego. Any tip to improve quality?

[pureexe]

I appreciate your great work!

However, I train Lego scene with both torch-ngp and instant-ngp at 20000 steps (200 epoch). instant-ngp has a sharper result (green rectangle) while torch-ngp is missing some part of the object (red rectangle). Any tip to improve the quality?

the training command is

OMP_NUM_THREADS=8 CUDA_VISIBLE_DEVICES=0 python train_nerf.py data/nerf_synthetic/lego --workspace trial_nerf_lego_tcnn --fp16 --tcnn --cuda_ray --bound 1 --scale 0.8 --mode blender

Best regards

[ashawkey]

You are right, the rendering quality is not still reproduced as mentioned in #17. Since there are lots of details in the original implementation, it may take some time for me to figure out the reason.
Currently, to improve quality (at the cost of speed), you can remove --cuda_ray, and enlarge --num_steps and --upsample_steps to use the pytorch ray marching.

[ashawkey]

The real reason is instant-ngp use all train/val/test set to train, which is set to default in GUI mode now. You can try it now!

[Dengzhi-USTC]

Hi, have you solve the problem about the ray marching?