Models and Editorial Workflows
Last updated: November 9, 2026
Unipax Science Studio includes an internal science-content studio and a video-clipping tool. Their data flows and retention rules differ.
1. Science studio
The studio is for authorized members. Administrators configure external providers, compatible gateways, model names and voices in Settings. A service marked unconfigured is not called. Saving a configuration or listing models does not verify media-generation capability.
| Stage | Integration | Content sent to the configured service |
|---|---|---|
| Article and storyboard assistance | Claude-compatible API | Editorial questions and angles, source descriptions, verified claims and writing instructions |
| Scene images | Image-generation API | Shot descriptions and style instructions |
| Image-to-video | Grok-compatible API | Selected reference images, motion instructions and duration |
| Speech synthesis | Speech API | Confirmed shot narration and voice settings |
| Captions and rendering | FFmpeg on our own servers | Selected assets, narration, captions and timeline |
The saved configuration and task records identify the actual provider, gateway and model. When a compatible gateway is configured, the listed content is sent to that gateway and processed under its service arrangements. Credentials are decrypted only on the server and are not displayed to ordinary members.
Model output is an unreviewed draft or asset. Models cannot declare sources, claims or work approved. Knowledge, scripts, visuals and final output have separate revision-bound review records. Editing content invalidates the applicability of prior approvals to the new revision.
Scientific structures, arrows, numbers and relationships require checkable diagrams. Generated scenes are illustrations, not experimental evidence. Automated checks do not replace scientific review.
The initial publishing workflow exports article drafts, assets and video packages. Operators publish through official platforms and record article and video results separately. The studio does not publish externally on its own.
2. Video clipping
The clipping workflow preserves source footage and audio; it does not generate new people or narration.
| Purpose | Model or tool | Runtime |
|---|---|---|
| Chinese and other speech recognition | SenseVoice-Small | Our processing machine |
| Other-language recognition and fallback | Whisper / faster-whisper | Our processing machine |
| Voice activity detection | FSMN-VAD | Our processing machine |
| Highlight selection and titles | DeepSeek, according to production configuration | External service, transcript text only |
| Person tracking and reframing | YOLOv8, MediaPipe | Our processing machine |
| Shot detection | TransNetV2, PySceneDetect | Our processing machine |
| Captions and encoding | ASS / FFmpeg | Our processing machine |
The clipping tool transcribes and processes video locally. Its highlight-selection text model does not receive the original audio or video. Transcripts and titles pass the clipping tool’s moderation process; see the Usage Policy.
3. Provenance and corrections
Creators must check asset permissions, scientific explanations, references and applicable platform disclosure requirements. Export packages retain AI-use notes, source lists and revision records; operators must also complete the publishing platform’s own declarations. Report content problems to support@unipax.cn.