Hyper3D / Deemos / Media2Face / GNPFA
Research diffusion model that animates a 3D face from speech with text or image style guidance.
Provider-documented facts. “Not collected” means we have not verified that field yet, not that the feature is unsupported.
Paper submission and revision dates are not weight-release dates.
Frame windows are not text-token context limits; overlapping windows handle longer audio subject to memory.
Hardware-specific reported throughput, not a hosted endpoint SLA.
No current hosted API price or callable checkpoint selector verified.
Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.
Not collectedNo public native selector for this research edition verified.
Endpoint pricing has not been collected. Check the provider for applicable rates.
No commercial retention or training-use contract assigned to a research publication.
Caching can reuse earlier input processing. Availability, lifetime and billing depend on the endpoint. A cache hit still generates a new response.
The system processes eligible input and retains reusable state.
Reuse depends on matching input, session state and the cache policy.
Reusing state can reduce repeated work. Any charges follow the billing rules below.
No model-specific caching contract verified.
No model-specific caching contract verified.
No model-specific caching contract verified.
No model-specific caching contract verified.
No model-specific caching contract verified.
No model-specific caching contract verified.
No model-specific caching contract verified.
No model-specific caching contract verified.
No verified cache rate for this endpoint. An empty rate is not a zero charge.
Caching reviewed 14 Oct 2026Paper Table 2, full model. Two-hour M2F-D holdout with similar composition to training data. Lower is better. Maximum L2 error over lip vertices per frame. Not an independently reproduced benchmark. Baselines without head pose omit BA; differing facial topology prevents vertex metrics for SadTalker.
Evaluation source (opens in a new tab)Paper Table 2, full model. Two-hour M2F-D holdout with similar composition to training data. Lower is better. Deviation of upper-face motion standard deviations. Not an independently reproduced benchmark. Baselines without head pose omit BA; differing facial topology prevents vertex metrics for SadTalker.
Evaluation source (opens in a new tab)Paper Table 2, full model. Two-hour M2F-D holdout with similar composition to training data. Higher is better. Audio/head-motion beat synchronization. Not an independently reproduced benchmark. Baselines without head pose omit BA; differing facial topology prevents vertex metrics for SadTalker.
Evaluation source (opens in a new tab)Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.
Qualified unreleased-weight research classification: the official project labels code and data upcoming, and its links lead back to the project. No original weight release was established in reviewed sources. This is not a claim of permanent proprietary licensing; publication alone does not release weights.
Classification evidence (opens in a new tab)