sbharadwajj / flare Goto Github PK

Thanks for open source such a great project！
When I trained the yufeng and marcel data sets, errors quickly occurred: Re-initializing main() because the training of light MLP diverged and all the values ​​are zero.
Code tested on RTX 3090Ti
How can I solve this problem?

Thanks for your great work, but it seems that I can't find a valid download-link in the Flare website

Training Deformer
STAGE 1: Using positional encoding (NeRF) for intrinsic materials
{'material_mlp_ch': 5, 'light_mlp_ch': 3, 'material_mlp_dims': [128, 128, 128, 128], 'light_mlp_dims': [64, 64]}

Re-initializing main() because the training of light MLP diverged and all the values are zero.

I got this error when running the code. Could you give some suggestions? Thanks very much.

Hi.
Thank you for great work.

While I was working on this work, I found out that the resulting uv map is corrupted.

I fount that although xatlas duplicates some vertices, current work does not apply this on the mesh class.

the resulting uv map is

however,

if I add two lines in the xatlas_uvmap a bit, like

@torch.no_grad()
def xatlas_uvmap(self):
    import numpy as np
    # Create uvs with xatlas
    v_pos = self.vertices.detach().cpu().numpy()
    t_pos_idx = self.indices.detach().cpu().numpy()
    vmapping, indices, uvs = xatlas.parametrize(v_pos, t_pos_idx)

    # Convert to tensors
    indices_int64 = indices.astype(np.uint64, casting='same_kind').view(np.int64)
    vmapping_int64 = vmapping.astype(np.uint64, casting='same_kind').view(np.int64)
    vmapping = torch.tensor(vmapping_int64, dtype=torch.int64, device=self.device)

    uvs = torch.tensor(uvs, dtype=torch.float32, device=self.device)
    faces = torch.tensor(indices_int64, dtype=torch.int64, device=self.device)

    self._uv_coords = uvs
    self._uv_idx = faces
    self.vmapping = vmapping

    # ADDED LINES
    self.vertices = self.vertices[vmapping]
    self.indices = faces

the resulting UV map goes

It may not affect the result, hope it helps.
especially about the tangent computation, which includes uv coordinates.

Can I get a mesh with color and texture?

Hi,
thank you for your great work!
I run the code on rtx 3090 and the training process works well. But the result I got looks so strange. Yesterday you told me to fix the tiny-cuda-nn warning tinycudann was built for lower compute capability ({cc}) than the system's ({system_compute_capability}). Performance may be suboptimal. and I just can't fix it by now. But maybe the result won't be that bad because of the warning?
Can you give me some advice to fix this. Thank you in advance!