FLOAT¶

Development line: project:float · thread float-public-releases
Last event: 2025-07-21 · 2 dated since 2024-12-05 · Researched: 2026-09-04 · confidence: medium

What it is¶

FLOAT is a PyTorch research implementation that animates a single, mostly frontal portrait from speech audio.

• Motion generation in a learned motion-latent space with flow matching.
• Lip, head, and expressive motion animation from audio.
• Emotion redirection and test-time head-motion editing.

The published package is inference-only, research and non-commercial, tested on Linux with A100 and V100 GPUs.

Use it for reproducible research demos, not as a commercially deployable talking-head stack.

Development line¶

• 2024-12-05 — FLOAT project page went public. FLOAT appeared as flow-matching audio-driven talking-portrait generation with an orthogonal motion latent space and speech-driven emotion.
• 2025-07-21 — FLOAT repository and supplementary resource links appeared. The official repository and manual checkpoint route pointed to that runnable inference path; we identified no new named model version.

What changed¶

FLOAT moved from a published method to a runnable inference package, then to a revised ICCV paper rather than a versioned product release.

• 2024-12-05: FLOAT was presented as flow-matching audio-driven talking-portrait generation with an orthogonal motion latent space and speech-driven emotion.
• 2025-02-17: the official repository released inference code and checkpoints under a non-commercial licence, making the method runnable.
• 2025-06-26: the project reported ICCV 2025 acceptance.
• 2025-07-21: the official repository and manual checkpoint route pointed to that runnable inference path; we identify no new named model version.
• 2025-09-19: arXiv v5 revised the paper; it does not state a new inference release.
• 2025-11-10: the main-branch history shows an update commit, but its functional delta is undocumented and GitHub lists no releases.

There is no tagged model-release train or published training code.

Treat the repository and its checkpoints as a research code drop, not a maintained release channel.

How to use this¶

As of 2025-07-21, consult the official FLOAT repository, its linked supplementary resource, and the project page to assess or reproduce FLOAT; verify exact artifact contents and usage terms separately.

1. Clone the official repository, create its Conda environment with Python 3.8.5, then install the documented CUDA 11.8 PyTorch build and requirements. — https://raw.githubusercontent.com/deepbrainai-research/float/main/README.md
2. Download the checkpoints with sh download_checkpoints.sh and place the main FLOAT, Wav2Vec2, and speech-emotion checkpoint assets in the documented checkpoint layout. — https://raw.githubusercontent.com/deepbrainai-research/float/main/README.md
3. Prepare a mostly frontal single-face portrait and a speech track; begin with automatic face cropping enabled. — https://raw.githubusercontent.com/deepbrainai-research/float/main/README.md
4. Run generate.py on a CUDA device with --ref_path, --aud_path, --seed 15, --a_cfg_scale 2, --e_cfg_scale 1, and --ckpt_path ./checkpoints/float.pth. — https://raw.githubusercontent.com/deepbrainai-research/float/main/README.md
5. For an explicit expression, add --emo with one of angry, disgust, fear, happy, neutral, sad, or surprise; adjust --e_cfg_scale only after a default run. — https://raw.githubusercontent.com/deepbrainai-research/float/main/README.md

Best practices¶

• Use a frontal portrait first. The model was trained on frontal head-pose distributions, and non-frontal inputs can reduce quality. — https://raw.githubusercontent.com/deepbrainai-research/float/main/README.md
• Keep automatic cropping on unless a measured comparison justifies --no_crop; disabling it can reduce performance, while the default crop can add black padding. — https://raw.githubusercontent.com/deepbrainai-research/float/main/README.md
• Extract vocals before inference when the audio has heavy background music; the authors point to ClearVoice for this preparation. — https://raw.githubusercontent.com/deepbrainai-research/float/main/README.md
• Start emotion guidance at the documented default of 1; the authors suggest 5–10 only when a stronger expression is required. — https://raw.githubusercontent.com/deepbrainai-research/float/main/README.md
• Keep work research-only and do not plan on fine-tuning from the official release: its licence is non-commercial and the authors state that training code will not be released. — https://raw.githubusercontent.com/deepbrainai-research/float/main/README.md

Superseded by this¶

• Before 2025-02-17: guidance that FLOAT was only a paper or project-page method and required waiting for official inference code and weights. Superseded by the official inference-code and checkpoint release.
• 2024-12-05: a method-only description is no longer sufficient setup guidance. Superseded by the official README's environment, checkpoint, preprocessing, and CLI instructions; the 2025-07-21 checkpoint link is a route to that runnable package, not evidence of a newly versioned model.

Still unknown¶

• The supplied Google Drive checkpoint URL could not be independently read, so its file contents, checksum, and whether it changed after 2025-07-21 remain unverified.
• No first-party Simplified-Chinese FLOAT documentation exists; third-party Chinese results were not used as evidence.
• The latest visible 2025-11-10 commit is only labelled “update”; no first-party changelog explains its functional effect.
• The official material does not state supported input duration, output resolution, VRAM use, or compatibility beyond its tested Linux A100/V100 setup.

Sources¶

Agent brief¶

• Subject: project:float, thread float-public-releases, 2 dated events 2024-12-05 → 2025-07-21.
• Practical note: As of 2025-07-21, practitioners should consult the official FLOAT repository and its linked supplementary resource in addition to the project page when assessing or reproducing FLOAT; verify the exact artifact contents and usage terms separately.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.