You can deploy this model and start earning money today!

initialneil/
DEGAS

Views

No views yet

DEGAS: pretrained full-body Gaussian avatars

Trained avatars for DEGAS: Detailed Expressions on Full-Body Gaussian Avatars (3DV 2025), built on the DREAMS-AVATAR captures.

Code: github.com/initialneil/DEGAS
Dataset: initialneil/DREAMS-AVATAR
Paper: arXiv:2408.10588
Registration: our multiview SMPL-X tracker (the fitted result ships with the dataset)

Available avatars

Avatar	Trained on	Iterations	Face driven by
`P1_smplx`	P1C1	800k	the fitted SMPL-X expression + jaw
`P1_dpe`	P1C1	800k	a per-frame 512-d DPE code, mesh face neutralised
`P2_smplx`	P2C1	800k	the fitted SMPL-X expression + jaw
`P3_smplx`	P3C1	800k	the fitted SMPL-X expression + jaw
`P4_smplx`	P4C1	800k	the fitted SMPL-X expression + jaw

P1_smplx and P1_dpe are the same subject, the same data, the same architecture and the same schedule. They differ only in how the face is animated, so they are a clean side-by-side of the two options the code supports.

Each capture's C1 session is the training session and C2 is held out, so evaluating an avatar on PxC2 is a genuine cross-session drive.

There is no P5 or P6 avatar and there will not be one: P5C1 does not exist in the source captures and P6C1 was excluded, so those two subjects have nothing to train on. Their C2 sessions ship in the dataset purely as cross-reenactment drivers.

Held-out numbers

Two held-out settings, both at 800k, both on cam03 (the frontal tele closeup, which no avatar ever saw during training):

Avatar	unseen view (own `C1`, cam03)	unseen session (`C2`, cam03)
	PSNR / SSIM / LPIPS	PSNR / SSIM / LPIPS
`P1_smplx`	25.69 / 0.927 / 0.061	21.28 / 0.885 / 0.097
`P1_dpe`	25.60 / 0.926 / 0.062	21.06 / 0.884 / 0.101
`P2_smplx`	27.996 / 0.924 / 0.071	24.19 / 0.891 / 0.115
`P3_smplx`	23.16 / 0.770 / 0.111	21.18 / 0.767 / 0.173
`P4_smplx`	28.51 / 0.959 / 0.066	24.01 / 0.943 / 0.094

The unseen-session column is the honest generalisation number: the avatar is posed from a different recording session of the same person.

Do not read these as a ranking of the avatars. They are absolute, per-subject numbers on different people wearing different clothes at different distances from the camera, so a lower row is not a worse avatar. The one comparison that is controlled is P1_smplx vs P1_dpe, which share a subject, a schedule and an architecture, and even there the whole-image numbers barely separate them: a face is roughly 1% of the pixels, so these figures are dominated by torso and clothing. The difference between the two face paths only shows up in a mouth region derived from the SMPL-X jaw weights, where P1_smplx leads by about 0.58 dB.

Layout

Each avatar is stored exactly as degas_eval.py --model_path expects:

avatars/<NAME>/
  config.yaml                                    the run's own config, paths made portable
  avatar.json                                    provenance: subject, capture, face path
  point_cloud/iteration_800000/checkpoint.pt     the model (~1.3 GB)
  point_cloud/iteration_800000/point_cloud.ply   the Gaussians
  point_cloud/iteration_800000/smplx_refined.pt  the refined SMPL-X parameters

Usage

bash

git clone --recursive https://github.com/initialneil/DEGAS && cd DEGAS
# install per the repo's Setup section, and place SMPLX_NEUTRAL.npz under model/data/

pip install -U "huggingface_hub[cli]"
hf download initialneil/DEGAS --include "avatars/P1_smplx/*" --local-dir weights
hf download initialneil/DREAMS-AVATAR --repo-type dataset \
    --include "data/P1C2/*" --local-dir DREAMS-AVATAR

python degas_eval.py \
    --dat_dir "$PWD/DREAMS-AVATAR/data/P1C2" \
    --ip none \
    --model_path "$PWD/weights/avatars/P1_smplx" \
    --configs "$PWD/weights/avatars/P1_smplx/config.yaml" \
    dataset.cache_dir="$PWD/cache/P1C2_eval_cam3" \
    dataset.test.cam_select=[3] \
    "dataset.test.frm_list=np.arange(0, 293, 8).tolist()"

You pass the avatar's config yourself. degas_eval.py does not read the run's config.yaml behind your back; it used to append it last, which let a training-time value silently outrank the command line. Each avatar here ships a portable, fully resolved config.yaml (dat_dir: ???, no machine-specific paths), so naming it alone is enough. Anything required but missing is reported by name up front, not deep inside the model.

--model_path must be absolute. A relative one is resolved inside --dat_dir (os.path.join(dat_dir, model_path)), because that is where a training run writes by default, so weights/avatars/P1_smplx fails with a FileNotFoundError on a path you never typed. Hence $PWD.

Three more things fail quietly rather than loudly, so they are worth stating plainly:

Pass the avatar's own config.yaml, or it is not evaluated under the settings it was trained with. Command-line key=value overrides still outrank every config file, which is why the test split is set that way above.
Give every capture its own dataset.cache_dir. Decoded frames are named cam%02d/%08d.jpg with no capture in the path, so a shared cache would serve P1C1's frame 110 for P1C2's frame 110.
For P1_dpe, dataset.with_face_dpe must point at the session you are driving with, not the one it was trained on. These avatars store it as the relative value dpe, and the loader resolves relative paths against dat_dir, so they follow --dat_dir on their own. A config saved by your own training run holds an absolute path to the training session's codes instead, and there you must override it.

For P1_dpe, point --model_path and --configs at that avatar instead, and add dataset.with_face_dpe= for the session you are driving with.

Training setup

800k iterations per avatar, roughly 41 h on a single RTX 3090. Trained on all frames of the C1 session across 29 of 32 cameras at 2x (1024x750). cam03, the frontal tele face closeup, is held out as an unseen view, and the whole C2 session is held out as an unseen session. The DPE codes were extracted from cam07 and cam30, both frontal tele views inside the training split, so no evaluation view leaks into the face conditioning.

Which face path won

On P1, with identical data, schedule and architecture, P1_smplx beat P1_dpe. That is why the other released avatars use the SMPL-X path. P1_dpe is the paper-faithful formulation, and the right choice when you have no trustworthy face fit, but it is not the stronger one on this data.

Read that carefully, though: whole-image and even head-crop metrics could not separate the two at all, differing only in the fourth decimal. A face ablation moves roughly 1% of the pixels, so whole-image PSNR is dominated by torso and clothing. Only a mouth region defined from the jaw-driven SMPL-X vertices distinguished them (PSNR +0.58, SSIM +0.015, LPIPS -10%). It is one subject, so treat it as a direction, not a settled result.

Limitations

P1_dpe evaluation is not deterministic. The per-frame face code is a random convex combination of the two camera codes, redrawn on every sample, and that path runs at eval time as well as during training. Restrict the evaluation to one camera per frame if you need reproducible numbers.
One subject per avatar. These are personalised avatars, not a generalisable model.
Driving signal must be a registered SMPL-X sequence in the DREAMS-AVATAR convention.
Densification is disabled during training, and there is no oral-cavity geometry, so teeth render as a specular smear when the mouth opens wide.
The SMPL-X body model is not included here. Register at smpl-x.is.tue.mpg.de and install it yourself.

Licence

CC BY-NC-SA 4.0, non-commercial use only, matching the DEGAS code. The DREAMS-AVATAR captures these were trained on are CC BY-NC 4.0. Any commercial use needs formal permission first.

Citation

bibtex

@inproceedings{shao2025degas,
  title = {{DEGAS: Detailed Expressions on Full-Body Gaussian Avatars}},
  author = {Zhijing Shao and Duotun Wang and Qing-Yao Tian and Yao-Dong Yang and Hengyu Meng and Zeyu Cai and Bo Dong and Yu Zhang and Kang Zhang and Zeyu Wang},
  booktitle = {Proceedings of the International Conference on 3D Vision (3DV)},
  year = {2025}
}