spaceflow / config /default.yaml
multimodalart's picture
multimodalart HF Staff
SpaceFlow demo
43ca9e0 verified
Raw History Blame Contribute Delete
1.33 kB
# render
num_views: 0
# trellis
trellis_img_model_name: "JeffreyXiang/TRELLIS-image-large"
# Local pipeline config that mixes TRELLIS-image-large (encoder/decoder/SLAT
# decoders) with TRELLIS-text-xlarge (sparse-structure & SLAT flow models).
# Necessary for spatial-control: microsoft/TRELLIS-text-large does not ship a
# sparse_structure_encoder, but encode_spatial_control() requires one.
trellis_text_model_name: "config/trellis_pipeline"
flow_model_in_channels: 8
img_model:
rescale_t: 3.0
cfg_strength: 5.0
cfg_interval: [0.5, 1.0]
text_model:
rescale_t: 3.0
cfg_strength: 7.5
cfg_interval: [0.5, 9.5]
# self-similarity guidance
sim_guidance:
num_part_clusters: 30
steps: 300
learning_rate: 5e-4
loss_weight: 1.0
local_condition_dilation: 0
local_condition_min_cells: 192
local_condition_max_dilation: 1
# Multiply self-attention weights by this factor before the softmax when both
# voxels share a routed condition. 1.0 disables and keeps stock self-attention.
mixed_self_attn_boost: 2.5
# "legacy" preserves the original PartField/geometric router. "active_margin"
# lets active local SQs claim nearby occupied cells when the generated structure
# is slightly shifted relative to the input superquadric.
local_routing_mode: legacy
local_routing_margin: 2.0
log_every: 100