About "A German Shepherd" model generation effect

[songxia928]

I can generate "pineapple" model, and The effect is the same as in the paper, because the provided configuration is used.
but the "A german Shepherd" model I generated was particularly poor. Is there something wrong about my steps or config?

or how can I generate new models , and configuration parameters.

"A german Shepherd" model

geometry

appearance

my configs

I cant find the config file of "A german Shepherd" in configs, so I change the config of "dog" to "A german Shepherd", as fellow:
( ps. I only change the "text", "out_dir")
**geometry_A_German_Sheperd.json: **

  1 {
  2     "mode": "geometry_modeling",
  3     "sdf_init_shape": "custom_mesh",
  4     "base_mesh": "data/animal/animal.obj",
  5     "sdf_init_shape_scale": [1.2, 1.2, 1.2],
  6     "sdf_init_shape_rotate_x" : -45,
  7     "translation_y": -0.2,
  8     "random_textures": true,
  9     "iter": 3000,
 10     "coarse_iter": 1500,
 11     "save_interval": 100,
 12     "train_res": [512, 512],
 13     "batch": 6,
 14     "dmtet_grid" : 128,
 15     "display": [{"bsdf" : "normal"}],
 16     "train_background": "black",
 17     "validate" : true,
 18     "mesh_scale" : 2.5,
 19     "out_dir": "A_German_Sheperd",
 20     "text" : "A German Sheperd",
 21     "seed" : 42,
 22     "add_directional_text": false,
 23     "camera_random_jitter": 0.4,
 24     "fovy_range": [25.71, 45],
 25     "elevation_range": [-10, 45],
 26     "guidance_weight": 100,
 27     "early_time_step_range": [0.02, 0.5],
 28     "late_time_step_range": [0.02, 0.5]
 29 }

**appearance_A_German_Sheperd.json: **

  1 {
  2     "mode": "appearance_modeling",
  3     "base_mesh": "out/A_German_Sheperd/dmtet_mesh/mesh.obj",
  4     "random_textures": true,
  5     "iter": 1000,
  6     "coarse_iter": 300,
  7     "save_interval": 100,
  8     "texture_res": [ 2048, 2048 ],
  9     "train_res": [512, 512],
 10     "batch": 4,
 11     "kd_min" : [0.03, 0.03, 0.03],
 12     "kd_max" : [0.97, 0.97, 0.97],
 13     "ks_min" : [0, 0.08, 0],
 14     "ks_max" : [0, 0.9, 0],
 15     "display": [{"latlong" : true}, {"bsdf" : "kd"}, {"bsdf" : "ks"}, {"bsdf" : "normal"}],
 16     "envmap": "data/irrmaps/mud_road_puresky_4k.hdr",
 17     "env_scale" : 2.0,
 18     "train_background": "black",
 19     "validate" : true,
 20     "out_dir": "A_German_Sheperd_appearance",
 21     "text" : "A German Sheperd",
 22     "seed" : 42,
 23     "add_directional_text": true,
 24     "camera_random_jitter": 0.2,
 25     "fovy_range": [25.71, 45],
 26     "elevation_range": [-10, 45],
 27     "guidance_weight": 50,
 28     "sds_weight_strategy": 0,
 29     "early_time_step_range": [0.02, 0.98],
 30     "late_time_step_range": [0.02, 0.98]
 31 }

[ARuiChen]

{
    "mode": "geometry_modeling",
    "sdf_init_shape": "custom_mesh",
    "base_mesh": "data/animal/animal.obj",
    "sdf_init_shape_scale": [1.1, 1.1, 1.1],
    "sdf_init_shape_rotate_x" : 0,
    "translation_y": 0,
    "random_textures": true,
    "iter": 3000,
    "coarse_iter": 1500,
    "save_interval": 100,
    "train_res": [512, 512],
    "batch": 4,
    "dmtet_grid" : 256,
    "display": [{"bsdf" : "normal"}],
    "train_background": "black",
    "validate" : true,
    "mesh_scale" : 2.6,
    "out_dir": "Sheperd",
    "text" : "A standing German Sheperd",
    "seed" : 42, 
    "add_directional_text": true,
    "camera_random_jitter": 0.45,
    "fovy_range": [25.71, 45],
    "elevation_range": [-10, 50],
    "guidance_weight": 50,
    "early_time_step_range": [0.05, 0.5],
    "late_time_step_range": [0.05, 0.5],
    "if_flip_the_normal": true
}

See Google Drive to visualize the training process.
Test on 8 gpus. 4 gpus should also work. Single gpu may not work because I have not optimized it specifically, but you can have a try.
Some experience in parameter adjustment: prompt words should include the accurate state of the target, such as standing, jumping, etc; When you find that the results are not smooth, reduce the guidance_weight; Try adjusting the elevation_range; When the generated object has multiple faces, add 'if_flip_the_normal': true; Set appropriate object proportions; Set appropriate object posture.

sheperd_geometry.mp4

[songxia928]

Thanks :D