"""Offline generation to a file. python generate.py --sec 4 --out motion.npz "walk" "Wave" "run/jog" "jump" Each prompt is used for --sec seconds, in order, in ONE continuous motion (the context is never reset). Output .npz (30 fps): positions (T, 22, 3) world joint positions in meters (y up), SMPL joint order rotations (T, 22, 3, 3) world joint rotations root_xz (T, 2) pelvis path on the floor prompt (T,) prompt used for each frame parents (22,) skeleton parents """ import os, sys, argparse import numpy as np, torch sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "live")) import live_server as L # loads the models (GPU) def main(): ap = argparse.ArgumentParser() ap.add_argument("prompts", nargs="+") ap.add_argument("--sec", type=float, default=6.0, help="seconds per prompt") ap.add_argument("--s", type=float, default=4.0, help="text strength (classifier-free guidance)") ap.add_argument("--seed", type=int, default=1) ap.add_argument("--out", default="motion.npz") a = ap.parse_args() L.ENG = L.Engine() torch.manual_seed(a.seed) w = L.reset_world() steps = max(1, int(round(a.sec * 30 / 4))) P, R, T, txt = [], [], [], [] for p in a.prompts: cond = L.embed(L.tok, L.clip, [p]).float() for _ in range(steps): for f in L.one_step(w, cond, a.s): rel = np.array(f["p"], np.float32).reshape(22, 3) rel[:, 0] += f["tr"][0]; rel[:, 2] += f["tr"][1] P.append(rel); R.append(np.array(f["r"], np.float32).reshape(22, 3, 3)); T.append(f["tr"]); txt.append(p) np.savez_compressed(a.out, positions=np.stack(P), rotations=np.stack(R), root_xz=np.array(T, np.float32), prompt=np.array(txt), parents=np.array(L.B.PARENTS), fps=30) print("saved %s: %d frames (%.1f s)" % (a.out, len(P), len(P) / 30)) if __name__ == "__main__": main()