Question about parameters

[masonintokyo]

I noticed that the texture quality of my output did not meet my expectations, especially when compared to the demo results which looked significantly better.

Could you please help me understand which parameters might influence the texture quality? Additionally, I would like to know the specific parameters used in the demo for comparison.

Here are the default parameters I used:

model_config:
  target: src.models.lrm_mesh.InstantMesh
  params:
    encoder_feat_dim: 768
    encoder_freeze: false
    encoder_model_name: facebook/dino-vitb16
    transformer_dim: 1024
    transformer_layers: 16
    transformer_heads: 16
    triplane_low_res: 32
    triplane_high_res: 64
    triplane_dim: 80
    rendering_samples_per_ray: 64
    grid_res: 128
    grid_scale: 2.1


infer_config:
  unet_path: ckpts/diffusion_pytorch_model.bin
  model_path: ckpts/instant_mesh_large.ckpt
  texture_resolution: 1024
  render_resolution: 1024

Input image

Result

Any guidance or suggestions would be greatly appreciated!

Thank you for your time and assistance.

[bluestyle97]

It's indeed due to the unsatisfying human body/face generation capability of Zero123++, which leads to distortions and multi-view inconsistency. Please try using front-view images as input.

[PierreBEGOUT]

I know the issue is answered but i'm also interested to know which paramters can I change to have a better quality meshes !

Thank you for your time and assistance

[sumanttyagi]

@bluestyle97 iso-surface extraction resolution is 256 with higher value would help here ?

[PierreBEGOUT]

@sumanttyagi And where is this parameter, I can't find it anywhere.