MissawB/procedural-quadrupeds

Procedural quadrupeds

500 synthetic quadrupeds, each given as a 35-joint kinematic skeleton and a 2048-point surface sampled around it. Five mammal families, 100 shapes each: felidae, canidae, equidae, bovidae and hippopotamidae.

It is the training set of creature-flow-point-dit, a flow matching model that generates the surface from the skeleton. You can see generated shapes in the viewer Space.

Load

from datasets import load_dataset
import numpy as np

ds = load_dataset("MissawB/procedural-quadrupeds", split="train")
row = ds[0]
points = np.array(row["points"], dtype=np.float32)  # (2048, 3)
joints = np.array(row["joints"], dtype=np.float32)  # (35, 3)
print(row["family"], points.shape, joints.shape)

The generator and the training code are on GitHub.

Fields

Field	Type	Content
family	string	One of felidae, canidae, equidae, bovidae, hippopotamidae
family_id	int8	Index of family in that list, 0 to 4
points	2048 × 3 floats	Surface points, y up
joints	35 × 3 floats	Joint positions in the same frame as points

The skeleton topology is the same for every row. Joint names in order:

pelvis, spine_1, spine_2, spine_3, neck_base, head, jaw, then for each front leg (l_, then r_) collar, shoulder, elbow, wrist, front_paw, front_toe_1, front_toe_2, then for each back leg (l_, then r_) hip, knee, ankle, back_paw, back_toe_1, back_toe_2, and finally tail_1, tail_2.

Parent of each joint (-1 for the root):

[-1, 0, 1, 2, 3, 4, 5, 3, 7, 8, 9, 10, 11, 11, 3, 14, 15, 16, 17, 18, 18,
 0, 21, 22, 23, 24, 24, 0, 27, 28, 29, 30, 30, 0, 33]

How it was made

Each shape comes from a parametric generator (data/smal_quadrupeds.py in the model repository), generator version 4, seed 42. Every family has base proportions for spine, legs, neck, head, tail and body thickness, plus a neck elevation (20° for bovids, 49° for the others). Each instance jitters those proportions, sets the leg stance in proportion to body thickness, poses the skeleton so all four feet rest on the same ground plane, and samples points on cylinders and ellipsoids around the bones.

Stance and neck elevation were added in version 4 so that body thickness and head carriage show in the skeleton itself. A model conditioned only on the skeleton could not tell bovids from horses without them.

Limitations

• Not real animals. The topology is inspired by SMAL, but the surfaces are simple primitives, not scans or the SMAL body model.
• Little variation within a family. Proportions vary by about ±3 %, so a family is close to a single template.
• One neutral standing pose for every shape.