Product frame
Short-video generation looks like media automation, but the real work is workflow design: each intermediate state has to be understandable for the system to keep improving.
Exploration · Local AI video factory
HLN Machine explores a local AI video factory. It is not about one generated result, but about splitting generation into a white-box pipeline that can be inspected, rerun, and repaired.
Narrative
Short-video generation looks like media automation, but the real work is workflow design: each intermediate state has to be understandable for the system to keep improving.
The pipeline is organized around IR.json, checkpoints, dependency hashes, partial reruns, and human-reviewable intermediate states, so failures can be located instead of hidden.
This is a lab for AI engineering judgement: transparent, diagnosable, and recoverable rather than packaged as a finished product.
System
| Component | Type | Visibility | Role |
|---|---|---|---|
| Seed intake | Input stage | Internal | Turns news or ideas into processable video seeds. |
| Script / subtitle pipeline | Generation stage | Internal | Generates reviewable short-video text structure. |
| B-Roll assembly | Media stage | Internal | Connects script structure to visual material. |
| Checkpoint system | Observability | Internal | Supports restart, diagnosis, and partial repair. |
Public URLs
Public exploration notes for the local AI video pipeline and white-box generation workflow.
www.notion.so/2c2c1b36c401806a909dd06baad8815a Media YouTube demoPublic video surface for seeing the local AI video pipeline and generated output in context.
www.youtube.com/watch?v=XnTwzDOpApc&t=11s