Not able to reproduce the paper results

[ahmadsalimi]

Hi. I've been trying to reproduce your method on the SPIn-NeRF dataset, but the results have noticeably lower quality than those presented in the paper. For example, here's the command I'm using to inpaint scene '1':

python DS_NeRF/run.py \
  --config DS_NeRF/config/config_1.txt \
  --expname mvip-1 \
  --datadir data/1/ \
  --text "a stone park bench" \
  --text_depth "a gray scale image of park bench" \
  --text_normal "a normal map of park bench" \
  --is_colla_guidance \
  --colla_guidance_scale 7.5 \
  --normal_guidance_scale 7.5

However, the resulting NeRF renders look like this:

mvip-1_gt_images_009000_rgb.mp4

An example from another scene ('9'):

mvip-9_gt_images_009000_rgb.mp4

[chenhonghua]

Hi, one limitation of this work is that it requires tuning the parameters. I think you can try to change the normal SDS scale to first check if the depth is well recovered, and then use RGB SDS to optimize the appearance.

[mldemox]

Thank you for your work, can you please provide specific visualization results? I need to compare it with your method.