KiMoDo

Kinematic motion-diffusion text-to-3D-motion model by NVIDIA. It is only 282M parameters, but memory is dominated by its Llama-3-8B text encoder: about 17 GB by default, dropping under 3 GB once that encoder is moved to CPU with TEXT_ENCODER_DEVICE=cpu. Fourteen cards carry a recipe here from a 12 GB floor, both Radeon boards included. NVIDIA Open License.