vibencode
UnstableML
A pipeline turning one photo and a short voice sample into an avatar that presents a product, with B-roll generated around it.
“A minute of finished product video costs us days, and none of that is the idea.”
An avatar built from one photo and ten seconds of audio, with B-roll and stills drawn from whichever model does that job best.
The cost was not in any one step. It was in the handoffs.
Scripting, filming, voiceover and editing each had a good tool, and each exported a file for the next to import with a person carrying it across. That work was invisible on every plan and took most of the calendar.
It also repeated. One idea becomes many cuts (different lengths, ratios, a different hook per platform), each a fresh pass over something already finished. What was needed was an orchestrator, with the generators as steps inside it.
Versioning did not make the work harder. It made it happen again.
- 01An avatar cloned from a single photograph and a ten-second voice sample, able to hold and wear the product it discusses.
- 02A scripted pass generating the piece to camera, lip-synced, so one avatar can carry a whole catalogue.
- 03B-roll and stills routed to whichever model is currently strongest for that shot, rather than to one house model.
- 04Per-platform variants cut from the same master, so a hook, a length and a ratio are settings rather than another edit.
- The avatarOne photo and ten seconds of audio, reusable across a catalogue.
- The model routerEach shot sent to whichever generator does that job best today.
- The variant setLengths, ratios and hooks cut from one master.
A pipeline that aggregates a dozen models inherits a dozen licences.
Several are unclear about commercial use of the output, and one bad inheritance contaminates everything downstream. Each model was audited before it went in, anything trained on unaccountable data was excluded, and the terms live in the pipeline rather than in a document somebody has to remember.
Generated output is uneven in a way that only averages out if nobody is watching closely.
A brand is exactly somebody watching closely, so every cut goes to a person before delivery. The role changed from assembling to judging. It did not go away, and the plan never pretended it would.
Running frontier models over every frame can cost more than the studio it replaces.
Open-source models wherever they hold up on this material, GPU scheduling rather than on-demand, and batching everything that need not be interactive.
A photo and a voice sample produce a presenter who can carry a catalogue, in the shapes each platform wants.
Because the router is not tied to one model, a better generator released next month is a configuration change rather than a rebuild.
- Booking a shoot per product
- Re-cutting finished work per platform
- Re-recording for each new script
This one sits across Agents and orchestration and Applied research. Most engagements touch two.
Is yours the same shape as this one?
Describe it in your own words. We will tell you where it lands.