How to make a PMV of one performer

Single-performer PMVs are one of the most requested kinds, and traditionally one of the most tedious: you scrub through everything you own, pull the clips featuring them, and only then start editing. On a large library that's an evening gone before the fun part.

There's a faster route. Show the AI about five clips of someone, and it finds the rest itself.

How it works, briefly

During ingest, Onset Engine analyses every clip with a vision model and stores a mathematical description of what's in it. Clips that look alike end up close together in that space.

So when you point at five clips and say "this is Subject X", it builds a profile from those examples and sweeps your entire library for anything close to it. No manual scrubbing — you're describing by example rather than searching by filename.

One thing to be clear about, because it sets expectations correctly: this isn't face recognition. It's general visual similarity. That has an upside and a limitation, and both are useful to know — more below.

Step by step

1. Find a few good examples

In the VIDEO tab, locate 4–6 clips of your performer. Quality of examples matters far more than quantity:

• Vary them. Different scenes, outfits, lighting, angles. Five near-identical clips teach a narrow profile that only matches more of the same.
• Keep them clean. Clips where they're clearly the subject, not sharing frame with three other people.
• Don't overthink it. You refine later; this is a starting point, not a final answer.

2. Create the Subject

Turn on Tag Mode, click your example clips to select them, then hit 🤖 Propagate Tag. Name it, and set a match threshold.

The threshold is a precision/recall dial. Higher means fewer but more certain matches; lower casts wider and drags in more false positives. The default is a reasonable starting point — you'll tune it after seeing results.

3. Review the borderline matches

Onset auto-tags the confident matches and puts the uncertain ones in a review queue: thumbnails, similarity percentages, and a ▶ button that plays the exact segment so you can check without leaving the dialog.

Tick the ones that are right, leave the rest. This is the step worth actually doing — every approval sharpens the profile, so the next sweep is better than the last.

4. Find more

Open the SUBJECTS tab, find your performer, hit 🔍 MORE. It re-scans using everything you've approved so far and returns a fresh review queue. Run it a couple of times and you'll surface clips the first pass missed.

5. Build the edit

Two options:

• Scope the bin. Hit FILTER on the Subject card to restrict the asset bin to that performer, then edit as usual.
• Reference them in a driver. Add "tags": ["@TheirName"] to a tier. On the peak tier, the drops are theirs. On every tier, the whole edit is. Start from the driver pack and edit the JSON — it's plain text.

Then pick a track, pick a preset, render. Full workflow in how to make a PMV.

Already using Stash?

Skip most of the above. Onset can import your performers straight from a local Stash library — every performer you've tagged arrives as a Subject with a profile already built, and Find More works immediately from there.

The limitation — and the trick hiding in it

Because this is visual similarity rather than face recognition, it keys on everything in the frame: the person, the framing, the setting, the pose. Which means:

The limitation: a performer in a radically different context — different hair, different lighting, very different shot type — may not match from your original examples. Fix: approve a few of those unusual clips through the review queue, and the profile widens to include them.

The trick: the same mechanism searches by pose or composition, not just by person. Seed it with five clips of a particular position or camera setup and it'll find that setup across your whole library, regardless of who's in them.

Searching by pose instead of by person

The process is identical — select examples, propagate, review — you're just choosing what the examples have in common. Instead of five clips of the same performer, use five clips of the same thing happening. Some ways that's genuinely useful:

• Position sequences. Build a Subject per position, then write a driver that moves through them as the song escalates. Each tier pulls from a different one, so the edit progresses instead of shuffling at random.
• Shot types. Seed on close-ups, or on wide shots, and you can hand the driver a consistent visual grammar — wides on the quiet parts, close-ups on the drops.
• Cutaways. Faces, reactions, hands, whatever you like to cut to. Tag the type once and every future edit has a pool to draw from.
• Look and setting. Outdoor clips, a particular room, a lighting style — anything visually consistent enough to describe by example.

Two practical notes. Keep the examples tight on the one thing they share: five clips of the same position but different performers, locations and lighting teaches "this position" cleanly, whereas five clips of the same performer in that position teaches a blurry mix of both. And name these Subjects so you can tell them apart later — pos_ or shot_ prefixes keep them from getting muddled with performer Subjects in the list.

For assembling a sequence that's visually coherent rather than a shuffle, this is arguably the more useful half of the feature — and it's the same three clicks.

Practical notes

• Ingest first. Only analysed clips can be matched — anything not ingested is invisible to this.
• Bigger libraries work better. More material means more for the sweep to find.
• Subjects are reusable. Build the profile once; every future edit can use it.
• It all stays local. Analysis and matching happen on your machine.