### Input # input rgba image path (default to None, can be load in GUI too) input: # input text prompt (default to None, can be input in GUI too) prompt: # input mesh for stage 2 (auto-search from stage 1 output path if None) mesh: # estimated elevation angle for input image elevation: 0 # reference image resolution ref_size: 256 # density thresh for mesh extraction density_thresh: 0.5 ### Output outdir: logs mesh_format: frames save_path: '' save_model: False ### Training # guidance loss weights (0 to disable) mvdream: False imagedream: False lambda_sd: 0 lambda_zero123: 1 # use stable-zero123 instead of zero123-xl stable_zero123: True lambda_svd: 0 # training batch size per iter batch_size: 14 # training iterations for stage 1 iters: 500 # training iterations for stage 2 iters_refine: 50 # training camera radius radius: 1.5 # training camera fovy fovy: 49.1 # align with zero123 rendering setting (ref: https://github.com/cvlab-columbia/zero123/blob/main/objaverse-rendering/scripts/blender_script.py#L61 # training camera min elevation min_ver: -30 # training camera max elevation max_ver: 30 # checkpoint to load for stage 1 (should be a ply file) load: # whether allow geom training in stage 2 train_geo: False # prob to invert background color during training (0 = always black, 1 = always white) invert_bg_prob: 0. n_views: 4 t_max: 0.5 ### GUI gui: False force_cuda_rast: False # GUI resolution H: 800 W: 800 ### Gaussian splatting optimize_gaussians: True position_lr_init: 0.001 position_lr_final: 0.00002 position_lr_delay_mult: 0.02 position_lr_max_steps: 500 feature_lr: 0.01 opacity_lr: 0.05 scaling_lr: 0.005 rotation_lr: 0.005 num_pts: 5000 sh_degree: 0 percent_dense: 0.1 density_start_iter: 3000 density_end_iter: 3000 densification_interval: 100 opacity_reset_interval: 700 densify_grad_threshold: 0.05 # deformation field deformation_lr_init: 0.00064 deformation_lr_final: 0.00064 deformation_lr_delay_mult: 0.01 grid_lr_init: 0.0064 grid_lr_final: 0.0064 ### Textured Mesh geom_lr: 0.0001 texture_lr: 0.2 deformation: net_width: 64 timebase_pe: 4 defor_depth: 1 posebase_pe: 10 scale_rotation_pe: 2 opacity_pe: 2 timenet_width: 64 timenet_output: 32 bounds: 1.6 plane_tv_weight: 0.0001 time_smoothness_weight: 0.01 l1_time_planes: 0.0001 kplanes_config: grid_dimensions: 2 input_coordinate_dim: 4 output_coordinate_dim: 32 resolution: [32, 32, 32, 12] multires: [1] no_grid: False no_mlp: False no_ds: False no_dr: False no_do: True use_res: True data_mode: svd downsample_rate: 1 # data_mode: c4d # downsample_rate: 2