ExAvatar¶

Development line: project:exavatar · thread exavatar-development
Last event: 2024-09-26 · 2 dated since 2024-08-01 · Researched: 2026-09-04 · confidence: medium

What it is¶

ExAvatar is a research pipeline for graphics and avatar practitioners who need body, hand, and facial-expression control from a single-person monocular phone video instead of X-Avatar’s 3D-scan/RGB-D input.

• SMPL-X and FLAME fitting: fits to the capture.
• 3D Gaussian avatar: trains from the fitted capture.
• View synthesis: renders novel body and hand poses, facial expressions, and viewpoints.

The project demonstrates a roughly 10-second neutral-pose phone scan. Unseen mouth interiors, palm surfaces, and dynamic clothing can fail. We use it for controlled offline avatar research when the SMPL-X fitting stack is acceptable. We do not use it where clothing deformation or unseen anatomy must be reliable.

Development line¶

• 2024-08-01 — ExAvatar project website shared. On 2024-08-01, ExAvatar was shared through its project website as a public project reference.
• 2024-09-26 — ExAvatar release repository shared. On 2024-09-26, ExAvatar published a dedicated GitHub release repository to provide a release-oriented source beyond the earlier project website.

What changed¶

2024-08-01 — ExAvatar appeared as an ECCV 2024 method combining SMPL-X with 3D Gaussian Splatting to make a controllable body, hand, and face avatar from a short phone scan.

2024-09-26 — ExAvatar_RELEASE made the official PyTorch reimplementation available with public assets and documented fitting, training, testing, and animation paths.

2026-09-04 — The official repository and documentation remain public, but GitHub lists no packaged releases or release notes. Reproducible use should record a chosen commit.

How to use this¶

As of 2024-09-26, consult the ExAvatar release repository for release-oriented project materials rather than relying only on the project website shared on 2024-08-01.

1. Create the Conda environment in an ExAvatar_RELEASE checkout: conda env create -f environment.yml, then conda activate exavatar. — https://github.com/mks0601/ExAvatar_RELEASE/blob/main/README.md
2. Install the fitting prerequisites: SMPL-X 1.1, FLAME, DECA, Hand4Whole, MMPose, Segment Anything, Depth Anything V2, and COLMAP; run the repository’s copy_code.py customizations. — https://github.com/mks0601/ExAvatar_RELEASE/blob/main/fitting/README.md
3. Put one person’s video at data/Custom/data/$SUBJECT_ID/video.mp4, extract frames, create all/train/test frame lists, set dataset = 'Custom', then fit SMPL-X with the documented camera mode. — https://github.com/mks0601/ExAvatar_RELEASE/blob/main/fitting/README.md
4. Generate foreground masks, create a background point cloud only for a static background, train with main/train.py, and inspect reconstruction output with test.py. — https://github.com/mks0601/ExAvatar_RELEASE/blob/main/avatar/README.md
5. For an avatar exported under output/model_dump/$SUBJECT_ID, supply SMPL-X driving parameters and run animation.py; use animate_view_rot.py for a rotating camera render. — https://github.com/mks0601/ExAvatar_RELEASE/blob/main/avatar/README.md

Best practices¶

• Capture outdoors or otherwise avoid strong illumination and hard shadows; choose the static- or dynamic-background preprocessing path accordingly. — https://github.com/mks0601/ExAvatar_RELEASE/blob/main/avatar/README.md
• Use fitting frames where most of the person is visible and untruncated; sampling about 5 fps from 30 fps is the documented speed-oriented starting point. — https://github.com/mks0601/ExAvatar_RELEASE/blob/main/fitting/README.md
• Use COLMAP for a moving camera; use the virtual-camera path for a static capture or SMPL-X parameter extraction. — https://github.com/mks0601/ExAvatar_RELEASE/blob/main/fitting/README.md
• If face fitting degrades, inspect flame_init/renders and verify that DECA results were prepared before rerunning the pipeline. — https://github.com/mks0601/ExAvatar_RELEASE/blob/main/fitting/README.md
• Keep the SMPL-X shape, joint-offset, and face-offset identity parameters identical between training and evaluation. — https://github.com/mks0601/ExAvatar_RELEASE/blob/main/avatar/README.md

Superseded by this¶

• 2024-09-26 — For implementation, the public ExAvatar_RELEASE repository and its fitting and avatar READMEs supersede pre-release or project-page-only guidance.

Still unknown¶

• The repository has no GitHub release package, and reviewed sources provide no current VRAM target, training-time expectation, or supported CUDA/driver matrix.
• No reviewed first-party source declares ExAvatar deprecated or names an official successor.
• No reproducible Chinese-language operating report or independent community hardware recipe was verified; the listed practices are maintainer guidance.
• The 2024-09-26 source permalink could not be retrieved and was not used as evidence.

Sources¶

Agent brief¶

• Subject: project:exavatar, thread exavatar-development, 2 dated events 2024-08-01 → 2024-09-26.
• Practical note: As of 2024-09-26, practitioners should consult the ExAvatar release repository for release-oriented project materials, rather than relying only on the project website shared on 2024-08-01.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.