RFD 1026 details: the weights table

Model id Parameters bf16 Q4_K_M Source
qwen_q4_k_m_image_edit 27.0 B 54.0 GB 14.85 GB published
krea2_turbo_text_to_image 16.9 B 33.8 GB 9.30 GB estimated
pixal3d_image_to_textured_mesh 12.02 B 24.05 GB 6.61 GB measured
seethrough_layer_decomposition 4.9 B 9.8 GB 2.70 GB estimated
trellis2_image_to_textured_mesh 4.0 B 8.0 GB 2.20 GB estimated
trellis_image_to_textured_mesh 1.8 B 3.6 GB 0.99 GB estimated
worldmirror2_reconstruct 1.2 B 2.4 GB 0.66 GB estimated
triposplat_image_to_splat 1.1 B 2.2 GB 0.61 GB estimated
skintokens_auto_rig 0.5 B 1.0 GB 0.28 GB estimated
p3sam_mesh_segmentation 0.4 B 0.8 GB 0.22 GB estimated
kimodo_text_to_motion 0.3 B 0.6 GB 0.17 GB estimated
unirig_auto_rig 0.125 B 0.25 GB 0.07 GB estimated
lingbot_map_environment_scan unknown unknown unknown unknown
voxhammer_text_mesh_editing 0 0 0 published
voxhammer_image_mesh_editing 0 0 0 published
weftspun_image_to_world 0 0 0 composite

VoxHammer carries no weights of its own. It edits with the TRELLIS.2 backbone, thus its cost is the TRELLIS.2 cost. weftspun_image_to_world runs TripoSplat and then TRELLIS.2. Its cost is the sum of those two. RFD 1033 covers the geometric algorithms, which hold no weights at all.

The blocklisted Hunyuan models add 9.2 GB. The shape model holds 3.3 B parameters, which is 6.6 GB. The paint model holds 1.3 B parameters, which is 2.6 GB.