search

FrozenBurning / Text2Light

[SIGGRAPH Asia 2022] Text2Light: Zero-Shot Text-Driven HDR Panorama Generation
https://frozenburning.github.io/projects/text2light/
Other
577 stars 46 forks source link

There are problems with the generated HDRIs. #6

Closed tdsuper closed 1 year ago

tdsuper commented 1 year ago

Dear authors, thanks for this amazing work!

I tried to generate the HDRIs using the pretrained models and render the balls using these HDRIs.

However, the rendering results seem to be brighter than yours. One of them is given below.

I don't know whether have I lose some steps, so I raise this issue.

hdr_ green grass field with trees and mountains in the distance _balls

FrozenBurning commented 1 year ago

Could you specify the engine you used for rendering this image? It seems that the appearance of renderings is washed out.

tdsuper commented 1 year ago

I render the image using the command line script with the same version blender, i.e. blender-3.1.2-linux-x64.

FrozenBurning commented 1 year ago

Got it. I've tested by my side. For this scene, you can tune the parameter for SR-iTMO to boost=5 and balance=2.7 at here. There is a scale ambiguity for the range of HDR map, the default parameter works well on most of the scene. These parameters take charge of fine-tuning the rendering quality.

tdsuper commented 1 year ago

I have re-generated the HDRI and then re-rendered the balls. The result is much similar to that you provided. Thanks!

But I still have another two questions. The first is whether you use the same parameters for the four rendering results at here. The second question is what is the intuitive meaning of these two parameters. How to adjust them for a given HDRI? Can they be set or adjusted autumatically given hdr_output at here?

FrozenBurning commented 1 year ago

The intuitive meaning of these parameters is to reduce the scale ambiguity when rendering, which offers another degree of flexibility for users to modify their lighting strength.