search

lzhnb / GS-IR

MIT License
276 stars 13 forks source link

Brigthness during relighting seems off in the garden scene #20

Open MrNeRF opened 1 month ago

MrNeRF commented 1 month ago

Hi, thanks for your wonderful work. I tried the garden scene and it seems that it generally appears to bright. Any idea why this is the case? 00000_snow 00000_sunset_240 00001_bridge

lzhnb commented 1 month ago

Changing the tone-mapping algorithm may solve this problem. In GS-IR, we roughly clamp the shaded value between 0 and 1, which is better to optimize. I recommend that you conduct other tone-mapping algorithms (e.g. ACES Film Curve) in the relighting for better results.

MrNeRF commented 1 month ago

I tried different algorithms. It basically does not change much of the appearance. Might be a little bit darker or brighter but in general it still looks way to bright. From your paper it is not clear to me if my results are deviating.

00000_snow 00002_snow 00003_snow 00004_snow envmap_relight

lzhnb commented 1 month ago

Do you enable the --gamma in training? It will use linear_to_sRGB for gamma correction.

lzhnb commented 1 month ago

The results in my paper are obtained by following the README.

MrNeRF commented 1 month ago

I followed the Readme. It only enables gamma here: python relight.py \ -m outputs/bicycle \ -s datasets/nerf_real_360/bicycle/ \ --checkpoint outputs/bicycle/chkpnt40000.pth \ --hdri datasets/TensoIR/Environment_Maps/high_res_envmaps_2k/bridge.hdr \ --eval \ --gamma

Where should I else enable it?

lzhnb commented 1 month ago

You should enable --gamma in both training and relighting.