Closed TLJParadox closed 3 years ago
Hello, In that experiment, we first reduced the model from 81k tris to 15k tris using the mesh reduction tool in Maya 2019, and used the reduced model as initial guess. The relevant parameters in the .json files are:
"texture_res" : [1024, 8192],
"train_res" : 2048,
"batch": 8,
"learning_rate" : 0.003,
"iter" : 5000,
"random_textures": true,"min_roughness" : 0.25
in this .json file ? "mtl_override": "data/dancer/dancer_diffuse.mtl"
?"random_textures" : false,
and no mtl_override
so the optimization will use the textures from the reduced initial guess as a starting point.
min_roughness
is not set in the json file, so will get the default value of 0.08
IIRC, the dancer models were generated in Simplygon as that produced better approximations than Maya. We haven't tested Unreal to generate LODs so cannot comment on the quality there.
could you tell me something about Automatic cleanup. in figure 7 this paper. What's your param settings(json) in Unreal Engine Paragon asset.