Reason for learned geometry being blury

[JinyanSu1]

Hi, Rui and Yongwei,
I find the geometry I learned is quite blury, do you think it is because of the parameter setting?

[ARuiChen]

Could you provide the configuration file and the number of GPUs used？

[ARuiChen]

"a full body photo of a girl wearing a mecha"
There are still some issues with avatar's face and hands, and I will optimize them accordingly.

girl.mp4

[JinyanSu1]

Thank you, Rui,
These are the configuration I used.
My assumption is that, it might not do well with anime styled characters. (This blury only happens when the input text is more of anime or cartoon style, especially when the 2D image has clear lines

{"mode": "geometry_modeling", "sdf_init_shape": "custom_mesh", "base_mesh": "data/person/person.obj", "sdf_init_shape_scale": [1.2, 1.2, 1.2], "translation_y": -0.1, "random_textures": false, "iter": 3000, "coarse_iter": 1000, "save_interval": 100, "train_res": [512, 512], "batch": 3, "dmtet_grid": 256, "display": [{"bsdf": "normal"}], "train_background": "black", "validate": true, "mesh_scale": 2.5, "out_dir": "Anime/Anime28", "text": "a girl, red long hair, single person, cute, black background, full-body shot", "seed": 42, "add_directional_text": true, "camera_random_jitter": 0.4, "fovy_range": [25.71, 45], "elevation_range": [-10, 45], "guidance_weight": 100}

[JinyanSu1]

Could you provide the configuration file and the number of GPUs used？

For GPU, I use 8 NVIDIA L4.

[ARuiChen]

It seems to be due to some bias towards female roles in stable diffusion 2.1. I can easily generate the 3D of Naruto, but it is more difficult to generate female characters such as Taylor Swift or a girl