video-diffusion-avatars (VDA): weights and examples

Video Diffusion Avatars: Taming Video Diffusion Models into Controllable 3D Avatars

Expression- and camera-controllable portrait animation on WAN2.2-Animate-14B. Code, installation and usage: https://github.com/SimonGiebenhain/video-diffusion-avatars. Project page: https://simongiebenhain.github.io/video-diffusion-avatars/.

The code downloads everything here with its own commands; you do not need to fetch files by hand:

python download.py weights     # the weights below, md5-checked (MD5SUMS.txt)
python download.py examples    # examples/ and examples/drivers/

Files

file	size	what it is
vda_full.safetensors	1.73 GB	the full model: VDA adapter for WAN2.2-Animate-14B
vda_fast_base.safetensors, vda_fast_student.safetensors	1.73 + 2.60 GB	the fast model: base adapter and the distilled 4-step causal student
xnemo_disentangler.pt	26 MB	expression disentangler for X-NeMo codes (used inside both models; also standalone)
text_emb_cache.pt	17 MB	cached prompt embeddings, so the 11 GB text encoder is not needed
driving_bank/nersemble8.npz, nersemble8.takes.json, nersemble8_canonical_body.npz	153 MB	the shipped driving sequence: 8 NeRSemble clips on one canonical person, 1029 frames
third_party/xnemo_motion_encoder.pth	247 MB	X-NeMo motion encoder (for your own driving videos), by the X-NeMo authors
third_party/pipnet/epoch59.pth	49 MB	PIPNet landmark model (preprocessing), mirrored unchanged
examples/<NAME>/	210 MB in total	25 precomputed example conditionings: reference image, proxy render, skeleton, expression codes
examples/drivers/*.mp4	9 MB	the example driving videos (NeRSemble performances)

The public WAN2.2-Animate-14B files are not part of this repository; python download.py base fetches them from their original source.

Licences

Our weights (vda_*.safetensors, xnemo_disentangler.pt, text_emb_cache.pt) are released under CC BY-NC-SA 4.0, like the code. They are adapters for WAN2.2-Animate-14B (Apache-2.0), which you download separately.

file	licence
third_party/xnemo_motion_encoder.pth	CC BY-NC 4.0, the X-NeMo authors
third_party/pipnet/	MIT, the PIPNet authors (see third_party/pipnet/LICENSE)
driving_bank/, examples/drivers/, examples/*/codes.npy and the driving motion in examples/	derived from NeRSemble: use is subject to the NeRSemble terms of use

Non-commercial use only. See https://github.com/SimonGiebenhain/video-diffusion-avatars for the full licence information and the citation.

Citation

@article{giebenhain2026vda,
  title   = {Video Diffusion Avatars: Taming Video Diffusion Models into Controllable 3D Avatars},
  author  = {Giebenhain, Simon and Kirschstein, Tobias and Schoneveld, Liam and
             Davoli, Davide and Chen, Zhe and Nie{\ss}ner, Matthias},
  journal = {arXiv preprint},
  year    = {2026}
}