PMV DRIVER PACK for Onset Engine
from thepmvguy.com
=================================

5 tuned drivers for PMV / HMV / cock hero / hypno editing:

  pmv_classic.json  - The canonical PMV arc, escalating by act:
                      tease -> foreplay -> sex -> climax.
                      Pair with: AGGRESSIVE or SENSUAL preset.
  hmv.json          - For animated source footage, escalating by act.
                      Caveat: I tune these against a live-action
                      library, so this driver's tier structure is
                      verified but its retrieval quality is not. If it
                      misses on your footage, see "Making your own".
                      Pair with: AMV preset.
  cock_hero.json    - Beat-game round structure with skewed tier
                      thresholds and a strong pull toward top-rated
                      clips at the peak.
                      Pair with: AGGRESSIVE preset, --pacing long-holds
                      or slow. Export the beat grid (the "BEATS" button
                      in the audio bin) for your beat bar overlay.
  hypno_loop.json   - Trance repetition: tiers share clips on purpose,
                      so the edit keeps circling a similar look. The one
                      driver here where tier overlap is the goal.
                      Pair with: HYPNOSIS preset, half-beat pacing.
  focus_rear.json   - A single-subject focus driver, escalating by what
                      is in frame (clothed -> underwear -> explicit).
                      Use it as the template for focusing on anything
                      else -- see "Making your own" below.

INSTALL
-------
Unzip these files anywhere you like - Documents is fine. There is no
special folder, and you do not need admin rights.

In Onset Engine, under Creative Direction -> Driver, click Browse and
pick one of the .json files. No restart needed.

Onset Engine loads a driver by its full path, so the file can live
wherever suits you. The drivers bundled with the app are in
C:\Program Files\Onset Engine\drivers\examples if you want to read them
for reference - but keep your own somewhere writable, so a reinstall
cannot wipe them.

TIPS
----
- Drivers pick WHICH clips play at each energy level. Presets control
  the LOOK (cuts, effects). Combine freely.
- Rate your best clips 4-5 stars. The peak tiers prefer highly-rated
  clips. It's a strong preference, not a hard rule - a lower-rated clip
  can still be chosen if nothing better fits, so you never get a gap.
- The descriptions are plain-text prompts - open any driver in a text
  editor and reword them for your library. It's just JSON.
- Tag your performers in the SUBJECTS tab, then add "tags": ["@Name"]
  to any tier to steer it toward a performer. It's a strong pull, not a
  lock - the tag is worth 30% of the score by default. Add
  "tag_weight": 0.6 to the driver's "global" block to push it harder.
- Full schema reference (every field and what it actually does to a
  clip's score):
    https://docs.onsetengine.com/driver-system/driver-masterclass/

MAKING YOUR OWN (the part that actually matters)
------------------------------------------------
Driver quality is almost entirely down to how you word the descriptions.
Both of these were measured against a real 1,200-clip library.

1. WRITE TAGS, NOT SENTENCES.
   The AI learned from image captions, so it responds to how people
   actually label pictures: short, blunt, concrete. Prose scores WORSE.
   Testing the same concept:

     "ass"                                <- strongest match
     "ass in thong"                       <- nearly as strong, more specific
     "rear view of a woman bending over"  <- noticeably weaker
     "close-up of a woman's buttocks"     <- weakest

   Use the words you'd type into a search box. Clinical vocabulary and
   full sentences both hurt.

2. ESCALATE BY WHAT'S IN FRAME, NOT BY SPEED.
   If your tiers describe the same thing at different speeds, they will
   play the same clips -- the AI reads content far better than pace.

   focus_rear.json originally escalated by pace ("gentle unhurried
   movement" -> "blurred fast motion") and its tiers shared up to 27% of
   their clips. Rewriting it to escalate by what's visible -- clothed,
   then underwear, then explicit -- dropped that to 4-13% on almost
   every pair. Same subject, same library, completely different result.

   Build the ladder from things you could point at in a still frame:
   clothing state, framing, act, toys. Not adverbs.

If you have Onset Engine installed from source, you can measure your own
wording before committing to it:

  python tools/test_prompts.py --driver mydriver.json

It reports, per prompt, how strongly it matches - and whether your tiers
actually pull different clips from each other.

thepmvguy.com/driver-pack/ for the full write-up.
thepmvguy.com/guides/ for workflows.
