beclab lares-media-create
Read this front door on any image, video, audio, or 3D generation / creation task. Load a reference only when that step is the one in front of you.
Trigger generation through Router. Router owns GPU scheduling. Calling FlowStudio, ComfyUI, or a local engine yourself starts a second GPU job and collides with whatever Router already launched.
Decision order
Stop at the first step that can do the job. Do not skip ahead.
- Router capability — does Router already offer this output family? If yes, call through Router. Load router.md.
- FlowStudio installed? — only if step 1 found nothing. Load flowstudio.md.
- Matching FlowStudio workflow? — only if FlowStudio is installed. A match still runs through Router (register / sync-models if the catalog is empty), never through FlowStudio HTTP.
- Other methods — only if Router has no row, FlowStudio is absent, or it has no matching workflow. Load fallback.md.
- Deliver — a successful produce is not finished until the file is in this conversation. Load deliver.md and run that landing before the reply. Do not stop at a URL, a job id, or a FlowStudio page.
Output families
Router mode is the catalog key. User words like “音频 / 声音” are not that key: speech and music are different modes. Probe --mode for the family below; an empty audio list is not a miss for a song.
| User wants | Router mode (step 1) | Also a hit |
|---|---|---|
| Image generate or edit | image_generation |
FlowStudio output=image |
| Video | video_generation |
FlowStudio output=video (may also appear under the FlowStudio image_generation provider) |
| Speech / TTS | audio with TTS flags |
router call speak |
| Music / song / generative audio | music_generation |
FlowStudio output=audio |
| 3D / mesh / glb | — | FlowStudio output=model3d |
Chat with vision is not generation. Transcribe is not TTS. ffmpeg_encode is transcode / test pattern, not a generative model.
Image, video, audio, and glb/gltf/obj preview under the reply after landing. Never reply with only a hyperlink.