3DTopia-XL / configs /inference_dit.yml
FrozenBurning
Update fast uv unwrap
cb029d0
raw
history blame
3.27 kB
debug: False
root_data_dir: ./runs
checkpoint_path:
global_seed: 42
inference:
input_dir:
ddim: 25
cfg: 6
seed: ${global_seed}
precision: fp16
export_glb: True
fast_unwrap: False
decimate: 100000
mc_resolution: 256
batch_size: 4096
remesh: False
image_height: 518
image_width: 518
model:
class_name: models.primsdf.PrimSDF
num_prims: 2048
dim_feat: 6
prim_shape: 8
init_scale: 0.05 # useless if auto_scale_init == True
sdf2alpha_var: 0.005
auto_scale_init: True
init_sampling: uniform
vae:
class_name: models.vae3d_dib.VAE
in_channels: ${model.dim_feat}
latent_channels: 1
out_channels: ${model.vae.in_channels}
down_channels: [32, 256]
mid_attention: True
up_channels: [256, 32]
layers_per_block: 2
gradient_checkpointing: False
vae_checkpoint_path:
conditioner:
class_name: models.conditioner.image.ImageConditioner
num_prims: ${model.num_prims}
dim_feat: ${model.dim_feat}
prim_shape: ${model.prim_shape}
sample_view: False
encoder_config:
class_name: models.conditioner.image_dinov2.Dinov2Wrapper
model_name: dinov2_vitb14_reg
freeze: True
generator:
class_name: models.dit_crossattn.DiT
seq_length: ${model.num_prims}
in_channels: 68 # equals to model.vae.latent_channels * latent_dim^3
condition_channels: 768
hidden_size: 1152
depth: 28
num_heads: 16
attn_proj_bias: True
cond_drop_prob: 0.1
gradient_checkpointing: False
latent_nf: 1.0
latent_mean: [ 0.0442, -0.0029, -0.0425, -0.0043, -0.4086, -0.2906, -0.7002, -0.0852, -0.4446, -0.6896, -0.7344, -0.3524, -0.5488, -0.4313, -1.1715, -0.0875, -0.6131, -0.3924, -0.7335, -0.3749, 0.4658, -0.0236, 0.8362, 0.3388, 0.0188, 0.5988, -0.1853, 1.1579, 0.6240, 0.0758, 0.9641, 0.6586, 0.6260, 0.2384, 0.7798, 0.8297, -0.6543, -0.4441, -1.3887, -0.0393, -0.9008, -0.8616, -1.7434, -0.1328, -0.8119, -0.8225, -1.8533, -0.0444, -1.0510, -0.5158, -1.1907, -0.5265, 0.2832, 0.6037, 0.5981, 0.5461, 0.4366, 0.4144, 0.7219, 0.5722, 0.5937, 0.5598, 0.9414, 0.7419, 0.2102, 0.3388, 0.4501, 0.5166]
latent_std: [0.0219, 0.3707, 0.3911, 0.3610, 0.7549, 0.7909, 0.9691, 0.9193, 0.8218, 0.9389, 1.1785, 1.0254, 0.6376, 0.6568, 0.7892, 0.8468, 0.8775, 0.7920, 0.9037, 0.9329, 0.9196, 1.1123, 1.3041, 1.0955, 1.2727, 1.6565, 1.8502, 1.7006, 0.8973, 1.0408, 1.2034, 1.2703, 1.0373, 1.0486, 1.0716, 0.9746, 0.7088, 0.8685, 1.0030, 0.9504, 1.0410, 1.3033, 1.5368, 1.4386, 0.6142, 0.6887, 0.9085, 0.9903, 1.0190, 0.9302, 1.0121, 0.9964, 1.1474, 1.2729, 1.4627, 1.1404, 1.3713, 1.6692, 1.8424, 1.5047, 1.1356, 1.2369, 1.3554, 1.1848, 1.1319, 1.0822, 1.1972, 0.9916]
diffusion:
timestep_respacing:
noise_schedule: squaredcos_cap_v2
diffusion_steps: 1000
parameterization: v
rm:
volradius: 10000.0
dt: 1.0
optimizer:
class_name: torch.optim.AdamW
lr: 0.0001
weight_decay: 0
scheduler:
class_name: dva.scheduler.CosineWarmupScheduler
warmup_iters: 3000
max_iters: 200000
train:
batch_size: 8
n_workers: 4
n_epochs: 1000
log_every_n_steps: 50
summary_every_n_steps: 10000
ckpt_every_n_steps: 10000
amp: False
precision: tf32
tag: 3dtopia-xl-sview
output_dir: ${root_data_dir}/inference/${tag}