Note Interactive viewer: watch 2048 noise points flow into a lion, horse, wolf, deer or chimera around its skeleton.

MissawB/creature-flow-point-dit

10.4M • Updated about 2 hours ago

Note 10.4M-parameter Point-DiT trained with conditional OT flow matching. EMA weights, code, 90% family recognition on unseen skeletons.

Note Training set: 500 procedural quadrupeds in 5 families, each a 35-joint skeleton with a 2048-point surface.

Note Flow Matching for Generative Modeling (Lipman et al., 2023): the conditional OT path used for training.

Note Scalable Diffusion Models with Transformers (DiT): source of the AdaLN-Zero blocks.