akashvenkateshwaran/txt2motion

KIT-ML Text-to-Motion (VAE + Denoiser)

Contents

• vae/ — skeleton-aware VAE (opt.txt + net_best_fid.tar)
• denoiser/ — latent diffusion denoiser, v-prediction, classifier-free guidance (opt.txt + net_best_fid.tar)
• eval_stats/ — mean.npy / std.npy used to normalize motion features for both the VAE and the denoiser (from the KIT-ML Comp_v6_KLD005 evaluator — not the same as dataset/KIT-ML/Mean.npy, which belongs to a separately trained VAE)

Citation

@inproceedings{hong2025salad,
  title     = {SALAD: Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing},
  author    = {Hong, Seokhyeon and Kim, Chaelin and Yoon, Serin and Nam, Junghyun and Cha, Sihun and Noh, Junyong},
  booktitle = {CVPR},
  year      = {2025}
}