OmniAvatar — Public release and demo availability¶

Development line: project:omniavatar · thread public-release-and-demo
Last event: 2025-08-05 · 2 dated since 2025-07-04 · Researched: 2026-09-04 · confidence: high

What it is¶

OmniAvatar is an open research system for talking avatar video generation from an image, audio, and text prompt.

• Lip-sync to speech audio.
• Facial expressions and body motion matched to audio.
• Text control over character action, background, and emotion.

The models are LoRA and audio-condition checkpoints over Wan2.1 T2V in 14B and 1.3B variants. Output resolution is limited to 480p. For local experiments, this is a working Wan2.1 stack. The public Space is third-party, so we do not rely on it as the only launch path.

Development line¶

• 2025-07-04 — OmniAvatar public project and source repository were linked. The official repository dates the release of these weights to 2025-07-02.
• 2025-08-05 — OmniAvatar Hugging Face Space was linked. On 2025-08-05, a Hugging Face Space named OmniAvatar appeared and referenced the earlier July project announcement. This adds a hosted demo surface, but the links alone do not establish its exact functionality, availability, or relation to a specific repository revision.

What changed¶

2025-07-04 — the public project included code, weights, and the 1.3B variant for Wan2.1. The official repository dates the release of these weights to 2025-07-02. 2025-08-05 — the third-party Hugging Face Space alexnasa/OmniAvatar appeared. On 2025-08-06, the OmniAvatar author team publicly confirmed the demo was successful, but it did not become an official inference API. 2025-06-23 — the OmniAvatar paper was published (arXiv:2506.18866). 2025-06-24 — the authors released inference code and initial weights.

How to use this¶

As of 2025-08-05, practitioners could locate OmniAvatar through its public repository and the linked Hugging Face Space; they should validate the current repository revision, Space behavior, and usage instructions before relying on either.

1. Clone the repository and install the pinned dependencies: PyTorch 2.4.0 with CUDA 12.4 and requirements.txt. — https://github.com/Omni-Avatar/OmniAvatar
2. Download compatible Wan2.1 T2V, the official OmniAvatar LoRA/audio-condition checkpoint, and wav2vec2-base-960h into the pretrained_models structure. — https://huggingface.co/OmniAvatar/OmniAvatar-14B
3. Prepare the input string formatted as [prompt]@@[img_path]@@[audio_path] and run scripts/inference.py with the 14B or 1.3B configuration. — https://github.com/Omni-Avatar/OmniAvatar
4. Open the Space for a brief preliminary check, but expect intermittent availability and replicate any critical result locally. — https://huggingface.co/spaces/alexnasa/OmniAvatar

Best practices¶

• Start with 480p and prompt/audio CFG in the 4–6 range; increase audio CFG if lip-sync is not stable enough. — https://github.com/Omni-Avatar/OmniAvatar
• Compose the prompt from the first-frame description, person behavior, and background as needed. — https://github.com/Omni-Avatar/OmniAvatar
• Plan VRAM ahead for 14B: the authors report 36 GB without memory management, 21 GB with 7B persistent parameters, and 8 GB with zero persistent parameters on a single GPU. — https://github.com/Omni-Avatar/OmniAvatar
• Do not use the Space as a production endpoint: user discussions report blank screens and issues during local cloning; run official code when reproducibility is required. — https://huggingface.co/spaces/alexnasa/OmniAvatar/discussions/1

Superseded by this¶

• 2025-06-24 — the paper-only or examples-only status is obsolete: the authors released inference code and weights.
• 2025-07-02 — the 14B-only status is obsolete: the authors released OmniAvatar weights for Wan2.1 T2V-1.3B.

Still unknown¶

• The response schema does not contain event_findings and new_events fields; their confirmed facts sit in what_changed rather than separate structured objects.
• Space SLA and stable interactive uptime remain unconfirmed: the page shows Running on Zero, and later user messages report a blank screen.
• The official repository sets no production API, commercial support, or complete set of verified GPU configurations.

Sources¶

Agent brief¶

• Subject: project:omniavatar, thread public-release-and-demo, 2 dated events 2025-07-04 → 2025-08-05.
• Practical note: As of 2025-08-05, practitioners could locate OmniAvatar through its public repository and the linked Hugging Face Space; they should validate the current repository revision, Space behavior, and usage instructions before relying on either.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.