OpenMontage Is Now the Default Media Engine in Our AI Factory
For months, one of the biggest questions inside our AI Factory has been deceptively simple:
How do we stop treating B-roll as a bolt-on step and start treating visual production as an intelligent, inspectable system?
We now have an answer.
OpenMontage has completed its acceptance path and is now the default media engine in our AI Factory.
This was not a “we installed it and it launched” milestone. We pushed the system through real production tests, deliberately difficult visual requirements, failed generations, licensing checks, approval gates, mixed-media fallbacks, rendering problems, timing bugs, and independent validation.
By the time the final Organic Orchard proof passed, OpenMontage had produced a valid 90-second H.264 video at 1920×1080 with AAC audio, passed all six approved visual slots, preserved the original narration and SRT unchanged, and cleared 31 out of 31 tests.
The Old Problem: B-Roll Was Too Narrow
Traditional automated video workflows often follow a pattern like this:
That works until the narration asks for something the stock library does not contain. Then the workflow tends to settle for a vaguely related clip, repeat footage, or fail entirely.
Our goal was not simply to replace one stock-video downloader with another. We wanted the Factory to understand:
What should actually be visible on screen right now?
And then decide whether that should be stock video, archive footage, an open-license still image, a locally generated visual, a motion graphic, or another appropriate media type.
What We Built
The existing Visual Source Router and Pixabay Scout remain available, but only as legacy fallbacks for a genuine OpenMontage technical failure.
If OpenMontage simply cannot find a good stock clip, that is no longer considered a technical failure. It should try another media type.
The Organic Orchard Test Became the Real Challenge
One of our proof projects was a 90-second Organic Orchard segment. Several scenes were easy to find through stock footage. Others were not.
The system eventually reached one especially difficult visual requirement: an adult visibly biting a peach.
Not holding a peach. Not standing near a peach tree. Not smiling with a peach. Not bringing a peach toward the mouth. The person needed to be visibly biting the peach.
Our approval system refused to lower that standard.
The System Failed Correctly Before It Succeeded
Multiple generated attempts were rejected. Image-to-image preserved the composition but produced a peach floating awkwardly near the subject’s teeth. Localized inpainting damaged the mouth and fruit interaction badly enough that the results were clearly unusable.
Instead of approving something “close enough,” the workflow kept the slot blocked.
Then we tested another already-installed local SDXL-family checkpoint: CHEYENNE_v20.
One candidate finally passed. Seed 220503 showed the subject in profile with the peach physically at the bite point, the lips clearly interacting with the fruit, and the hand holding it in place. The other candidates were rejected because the fruit-and-mouth interaction was still ambiguous.
No cloud generation service was required. No paid image provider was activated. It was generated locally through our existing ComfyUI environment.
Mixed Media Became the Real Breakthrough
Once that final visual passed, the Organic Orchard manifest contained three approved locally generated stills, one approved open still, and two approved stock videos.
That is the point where the project stopped being about B-roll. This became a mixed-media production engine.
The Factory was no longer asking, “Which stock video should go here?” It was asking, “What media best communicates this part of the story?”
Then Rendering Exposed More Problems
The visual assets were finally correct, but the work was not finished.
The first mixed-media render failed because the adapter used the wrong animation schema. Then Remotion tried to treat a PNG image like a video. That exposed a renderer-selection problem: the video-only documentary renderer was being used for a mixed still/video project.
The Factory was adjusted to use OpenMontage’s native Explainer composition for mixed-media projects while retaining documentary-montage for video-only projects.
A 90-Second Video That Was Actually 16 Seconds
Then we found a more important bug. The render completed, and an early acceptance check even reported success, but ffprobe showed the file was only 16 seconds long.
That is exactly why we insisted on an independent Factory acceptance layer.
The problem was timeline semantics. The mixed-media renderer interpreted timing fields as output timeline positions while our Factory driver had supplied them like source-trim values. The overlapping cuts collapsed the output.
We corrected the timeline placement and strengthened acceptance so a playable MP4 is not enough: the artifact must also match the requested duration.
Then It Rendered 91 Seconds
The corrected video rendered at 91 seconds instead of 90. Again, the validator rejected it.
The extra second came from OpenMontage’s native final fade. Rather than loosen the acceptance test, the Factory reserved the final second inside the requested duration.
The result: exactly 90.000000 seconds.
Final Acceptance
The completed Organic Orchard proof passed the entire local OpenMontage acceptance path:
- Factory adapter invoked the real local OpenMontage runtime.
- Canonical narration and SRT were accepted without modification.
- All six approved media slots passed the manifest gate.
- A local 90-second video was produced.
- Checkpoint, invocation, provider, and error evidence remained inspectable.
- Independent Factory acceptance returned accepted: true.
- ffprobe verified H.264, 1920×1080, AAC, and exactly 90 seconds.
- Representative frames confirmed the peach bite and bee pollination visuals.
- The Coolidge proof independently passed regression.
- 31/31 tests passed.
- Canonical input hashes remained unchanged.
That produced the milestone we had been working toward: READY TO MIGRATE B-ROLL ROUTING.
And Then We Flipped the Route
The next production smoke test was not another specially tuned acceptance project. The Factory’s normal media routing was changed so OpenMontage became the default engine, while the old Visual Source Router, Pixabay Scout, and generic Shot Builder retrieval path became explicit legacy fallbacks.
A separate project, bridging-intent-gap-goal-oriented-ai, then entered through the normal Factory route. OpenMontage was selected automatically, legacy fallback was not used, independent acceptance passed, and the resulting 90-second video was H.264 1920×1080 with AAC audio.
That means this is no longer just an experiment. OpenMontage is the default media route.
What OpenMontage Replaces
We did not delete every older Factory component. Instead, OpenMontage now takes over the primary role previously spread across several systems:
- visual planning
- open-stock discovery
- media ranking
- still-image fallback
- local generated-image fallback
- mixed-media composition
- approval-aware production
- technical evidence
- rendering
The legacy Visual Source Router and Pixabay Scout remain available only for a proven OpenMontage runtime or technical failure.
What OpenMontage Does Not Control
We deliberately kept several boundaries intact. The Factory still controls canonical narration, canonical SRT, project-state authority, final acceptance, thumbnails, editorial approval, and publishing to Blogger, Facebook, and YouTube.
OpenMontage is the media engine. It is not the entire company.
Local AI Generation Is Now Part of the Media Ladder
One of the strongest outcomes was proving that ComfyUI can become a native fallback inside the media pipeline instead of a separate workflow.
Generated assets still face the same approval rules as everything else. They need provenance, technical validation, relevance, visible scene compliance, and explicit approval.
Why the Approval Gate Matters
Several images were technically successful generations. They loaded. They had the correct dimensions. They showed a person and a peach.
But they were wrong.
The workflow rejected them.
A production system needs to be able to say:
This completed successfully, but it is not good enough.
That distinction is essential if AI automation is going to move beyond toy workflows.
The Factory Is Getting Smaller — and Smarter
One of our broader goals has always been to avoid maintaining custom code simply because we already wrote it. If a stronger open-source component can own a capability more effectively, the Factory should adopt it.
OpenMontage is beginning to eliminate the need for several separate custom media systems. That matters because adding a supervisory agent should make the Factory more capable, while OpenMontage lets us retire lower-level plumbing.
The Next Phase: Hermes
OpenMontage is now the media engine. The next architectural phase is Hermes supervision.
Hermes should eventually sit above the Factory and decide whether a project is ready for media production, whether acceptance passed, whether a scene needs repair, whether a stage should be retried, whether human approval is required, and whether the project is ready for publishing.
From B-Roll Script to Media Engine
We started this migration trying to solve a B-roll problem.
We ended up building something considerably more useful.
The Factory now has a media-production layer capable of moving between real footage, open media, still imagery, locally generated visuals, approval gates, composition, and independent validation.
Most importantly, it can fail closed instead of silently pushing weak assets into production.
The Organic Orchard proof did not pass because we lowered the standard. It passed because the workflow kept working until every required scene satisfied the standard.
That is the difference between a collection of AI tools and an actual production system.
Comments
Post a Comment