jmunkberg
commented
1 year ago Hello @Williamarvin ,
With the dmtet file, are you referring to the surface mesh? That should be saved automatically after the first pass. Here is the relevant code: https://github.com/NVlabs/nvdiffrec/blob/main/train.py#L613
The dmtet tetrahedral mesh used in the optimization is read from disk as shown here: https://github.com/NVlabs/nvdiffrec/blob/main/geometry/dmtet.py#L171
There are instructions here if you want to generate tet meshes of different resolutions: https://github.com/NVlabs/nvdiffrec/tree/main/data/tets
From your parameters, "mesh_scale" : 7, looks fairly high. The tet grid should tightly cover the bounding box of your model for best results, otherwise triangles density is wasted. Easiest to look at the pngs from early optimization steps.
I would also recommend a batch size of 4-8 if you have enough memory.
After running train.py, I cannot seem to generate the dmtit file but instead, only output png files, do you know how i can do that? Also, it stops after all the training is done e.g 1000/1000, and showed psnr 7.89.. So my main concern is how to generate these dmtit file, not just png
Thanks in advance
These are the parameters { "ref_mesh": "data/fox/needed", "random_textures": true, "iter": 1000, "save_interval": 100, "texture_res": [ 1024, 1024 ], "train_res": [700, 700], "batch": 1, "learning_rate": [0.03, 0.03], "kd_min" : [0.03, 0.03, 0.03], "kd_max" : [0.8, 0.8, 0.8], "ks_min" : [0, 0.08, 0], "ks_max" : [0, 1.0, 1.0], "dmtet_grid" : 64, "mesh_scale" : 7, "camera_space_light" : true, "background" : "white", "display" : [{"bsdf":"kd"}, {"bsdf":"ks"}, {"bsdf" : "normal"}], "out_dir": "nerd_ehead" }