Models¶
OmniRT's models are owned by the registry: every model uses @register_model to declare its supported tasks, accepted adapters, minimum VRAM, and recommended preset.
From this stage on, model count is no longer the main goal. Maintenance is organized around the digital-human pipeline:
- Core: TTS, audio-driven avatars, realtime / resident workers, deployment, and benchmarks. These require real-hardware smoke evidence.
- Adjacent: avatar assets, backgrounds, idle video material, and post-processing for digital-human products.
- Experimental: existing general image / video integrations. They keep registry entries and basic tests, but do not promise dual-backend validation.
This section is organized around three tables:
| Page | Purpose |
|---|---|
| Supported Models | auto-generated full registry (equivalent to omnirt models) |
| Support Status | manually curated digital-human priorities, real-hardware smoke, and contraction status |
| Roadmap | digital-human main line, adjacent capabilities, and experimental model boundaries |
Query models from the CLI
omnirt models lists everything; omnirt models <id> dumps a model's ModelCapabilities (tasks, adapters, VRAM, recommended preset).