Log in
TechniqueBy HypernaturalVersion 3, updated

Match the source's push-in on a pinned last frame

hn/python-match-push-in

Crop a derived last frame so it is pushed in from your locked first frame exactly as far as the source shot's last frame is from its first, measured, so a first-and-last-frame render keeps the original camera move. Run after nano-banana-derive-last-frame; image models ignore 'push in 10%'.

SKILL.md

You give it the source shot's first and last frames, your locked first frame and your derived last frame; you get the last frame cropped to the source's push-in, at 1920×1080. Use it on every last frame made with nano-banana-derive-last-frame when the source shot moves in: the image model keeps the framing whatever the prompt says, and a pinned pair with no zoom makes the video lose the push-in.

Inputs

  • The source shot's first and last frames (stills at their exact frames).
  • Your locked first frame (1080p).
  • Your derived last frame: pass the 2K render (2752×1536), not a 1080p copy, so the crop keeps resolution.

Run it

Write references/push_in.md's script into the job's scripts/ as push_in.py, unchanged, and run it with exec python:

exec python scripts/push_in.py --source-first <job>/shots/shot09_first.jpg --source-last <job>/shots/shot09_last.jpg \
    --first <job>/stills/shot09_first_p.png --last <job>/stills/shot09_last_n.png --out <job>/stills/shot09_last_r.png

It measures the source's move (zoom about the centre plus shift, searched from 0.80 to 1.25 in steps of 0.005), measures how your last frame already sits against your first, crops the difference out of the 2K render, and conforms the crop to 1080p. It tries both signs of the shift and keeps the crop that lands closer.

What you get

JSON on stdout:

  • target: the source's last frame against its first. zoom below 1 is a push-in (0.94 is about 6% in); shift_px is the drift at 1080p; match is the correlation after alignment (1.0 identical, a shot's own first against last ≈ 0.86).
  • ours_before: your uncropped last frame against your first (usually zoom ≈ 1.0).
  • got: the written crop against your first. It should equal target.
  • miss_px: |zoom error| × 1920 plus the shift errors in px.

Check it

  • got.zoom within 0.005 of target.zoom and the shift within a few px; miss_px in single digits to the low teens. Larger means the alignment locked onto the wrong thing (a big moving subject): check the frames, then re-run.
  • Put first and pushed-in last in the same option of a review: the set should grow evenly with nothing jumping.
  • The crop changes the scale of everything, small print included. Fix the logo after this step, on the cropped frame, with a match reference cut at the same scale (nano-banana-rerender-small-print, match-cut in python-crop-and-paste-back). A logo matched at a different scale changes width and slides across the prop mid-clip.

The full first-frame, last-frame, push-in, print sequence and its gates: guide-keep-small-details.

FAQs

Crop a derived last frame so it is pushed in from your locked first frame exactly as far as the source shot's last frame is from its first, measured, so a first-and-last-frame render keeps the original camera move. Run after nano-banana-derive-last-frame; image models ignore 'push in 10%'.

Install Hypernatural Desktop, a free Mac app, and connect it to Claude Code, Codex or Claude Desktop. Then describe the video you want and your agent finds the skill itself, or paste this page's link (https://hypernatural.ai/skill/hn/python-match-push-in) and it installs hn/python-match-push-in and follows it.

Nothing. The skill is free and runs with tools on your own Mac, like python.

Yes. Ask your agent to adjust it for your job, or to save its own version. You can publish your version under your own handle so anyone can install it.

Go make something cool.

Install the app, open your agent, describe the video. That's the whole onboarding.

Sticker HypernaturalSticker BannerSticker TVBoy