AvatarArtist¶

Development line: project:avatarartist · thread avatarartist-public-release-and-space
Last event: 2025-04-09 · 2 dated since 2025-03-27 · Researched: 2026-09-05 · confidence: medium

What it is¶

AvatarArtist is an open-domain portrait-to-4D-avatar pipeline for creators and researchers who need motion-driven output.

• portrait conversion: turns one portrait into a parametric 4D representation
• frame rendering: renders target frames from the reference image and driving signals
• pipeline access: includes released inference and input-processing paths

Development line¶

• 2025-03-27 — AvatarArtist launched its public project page and official repository. This established the local code-and-weights workflow.
• 2025-04-09 — AvatarArtist linked to a Hugging Face Space. No reviewed first-party release note proves a model revision on that date.

What changed¶

AvatarArtist moved from a local inference release to local and online demos. A training release is still not documented.

• 2025-03-26: the official repository announced inference code and pretrained models.
• 2025-03-27: the public project page and official repository established the local code-and-weights workflow.
• 2025-03-30: the official repository listed a Gradio demo release.
• 2025-04-02: the official repository listed an online-demo release.
• 2025-04-09: the Hugging Face Space appeared; no reviewed first-party release note proves a model revision on that date.
• 2026-09-04: the former ant-research GitHub URL redirected to robbyant-research; the repository still lists training code as pending.

How to use this¶

As of 2025-04-09, consult the AvatarArtist project page and repository, then test the linked Hugging Face Space for hosted evaluation.

1. Start with a portrait reference and a driving-motion source; AvatarArtist creates a 4D representation from the portrait and renders frames under the driving signals. — https://kumapowerliu.github.io/AvatarArtist/
2. Clone the repository, create the documented Conda environment with Python 3.9, and install requirements. — https://github.com/robbyant-research/AvatarArtist
3. Download the KumaPower/AvatarArtist weights into pretrained_model; obtain faceverse_v3_1.npy separately and copy it to both documented FaceVerse paths. — https://github.com/robbyant-research/AvatarArtist
4. Run python3 app.py with the demo data first; for controlled output, run inference.py with image, FaceVerse coefficients, motion inputs, output path, and optional --select_img. — https://github.com/robbyant-research/AvatarArtist
5. For custom media, use the official preprocessing path so the final dataset contains the required image, coefficient, motion, and render folders before inference. — https://github.com/robbyant-research/AvatarArtist
6. Use the browser Space only for a quick manual trial; its accessible page snapshot was marked Sleeping rather than an active endpoint. — https://huggingface.co/spaces/KumaPower/AvatarArtist

Best practices¶

• Begin with the supplied demo data and the documented preprocessing layout; only the final dataset directory is model input, while the other folders are intermediates. — https://github.com/robbyant-research/AvatarArtist
• Treat faceverse_v3_1.npy as required setup, not an optional extra: weights alone do not make the project runnable. — https://github.com/robbyant-research/AvatarArtist
• Keep work within released inference and data processing; the official to-do still lists training-code release as pending. — https://github.com/robbyant-research/AvatarArtist
• Use the redirected robbyant-research repository for new setup links; retain the ant-research URL only for historical references. — https://github.com/ant-research/AvatarArtist
• Do not treat the model card as hosted inference: it lists no Hugging Face Inference Provider deployment. — https://huggingface.co/KumaPower/AvatarArtist
• Treat the Space as optional exploration rather than a production dependency because the accessible snapshot showed it Sleeping. — https://huggingface.co/spaces/KumaPower/AvatarArtist

Superseded by this¶

• 2025-03-27: using https://github.com/ant-research/AvatarArtist as the canonical owner URL is superseded by its observed redirect to https://github.com/robbyant-research/AvatarArtist on 2026-09-04.

Still unknown¶

• The dated 2025-04-09 event text is unavailable, so we cannot tie it to the online demo beyond the linked Space.
• We did not run the Space end to end during review, leaving current health, cold-start behavior, and latency unverified.
• First-party sources give no supported GPU, VRAM, CUDA, or operating-system matrix.
• We found no community experience reports; these practices come from official requirements and distribution state.

Sources¶

Agent brief¶

• Subject: project:avatarartist, thread avatarartist-public-release-and-space, 2 dated events 2025-03-27 → 2025-04-09.
• Practical note: As of 2025-04-09, practitioners should consult the AvatarArtist project page and source repository, then use the linked Hugging Face Space as the hosted public access point when evaluating the project.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.