Privacy by Framing: Exclude at the Lens
TLDR
If your model needs human action, you need hands and motion, not identities. At posterior.xyz we frame privacy in before we press record, so the files you get were never carrying extra personal detail in the first place.
Blur after the fact leaves you with the risk
Redaction happens after the camera already saw too much. The sensitive frame was captured, stored, moved across disks, and opened by someone for review. You still pay for all of that.
We work the other way. You tell us what is in scope for the task. We write framing guides that hold that boundary on set. Camera position, lens choice, and set layout decide what can appear, and what cannot.
The payoff is practical. No blur artifacts for your model to learn around. No long redaction review where someone checks frame by frame that faces stayed covered. And a protocol your legal team can read before collection starts, not after.
Hands-only and face-out framing
For manipulation, assembly, cooking, tool use, and device interaction, faces rarely help the model. So we leave them out.
That looks like overhead, side-angle, or first-person views that crop above the shoulders. Workspace framing that keeps hands, tools, and objects centered while surroundings that could identify a person stay outside the shot. Clothing and accessory controls when you need a neutral appearance across sessions.
You still get HD to 4K video with contact detail intact. Grip, pressure, handoffs, spills, slips, the small corrections people make mid-task. The procedure stays. Identity does not.
This holds across kitchens, workshops, benches, and outdoor work areas. The set changes. The framing rule does not.
Audio discipline with scripted fictional content
Background speech leaks more personal detail than most teams expect. A passing conversation can drop names, places, and routines you never asked to collect.
Our fix is kinda boring on purpose. When a scenario needs speech, participants speak only scripted fictional lines written for the task and approved by you. Crew talk follows a quiet-set protocol during takes. When a scenario needs no speech, we collect without usable dialogue and we log that choice in the session notes.
You skip the scrub of hours of incidental audio. Transcripts match what you planned to collect, nothing extra.
Minimal technical metadata by design
Video files often carry more than pictures and sound. We keep the file headers down to what training and provenance actually need.
You get capture settings and session structure tied to the file. Device identifiers, participant details, and location traces stay out unless you ask for a specific field and sign off on how we handle it. Consent records live apart from training files, so you can show responsible sourcing without mixing personal data into your pipeline.
What this enables for you
Faster internal review, because the dataset was scoped at the source. Cleaner training signal, because the model sees hands, objects, and actions instead of identities to ignore. More time spent on coverage and performance, less on cleanup.
If you are collecting human activity in real consented spaces, and you need synced depth and motion or multimodal delivery around it, framing-first privacy keeps the project focused and defensible.
Start with a paid pilot. Both sides check fit before anything scales. Request a pilot and Talk to us.
