Instagram video remake

Remake a video the user is allowed to remake: replace ONE person with another (face, hair, body, wardrobe), keep everything else, optionally continue a shot past its end, and deliver a finished file with sound. Generation is MiniMax H3 in the user's own ComfyUI. Every step is a command of scripts/main.ts and is recorded in remake.json (schema ig.remake.v1) inside the run folder.

npx -y bun scripts/main.ts --help      # run from this skill's folder

Before anything: consent and scope

This edits a real person's likeness. Refuse to proceed, and say why, unless both hold:

• The person whose face is ADDED agreed to appear in this video. The user states it; init stores the statement (--consent). Do not invent it, and do not soften it into "it is public".
• The user may reuse the source video (their own, licensed, or a use they take responsibility for).

Never: remove clothing or sexualise beyond what the source shows, depict a minor, put words in someone's mouth, or imitate a public figure to mislead. A faithful recast of an existing scene is in scope; anything in that list is not, whatever consent is claimed.

The result is realistic AI video of a real person. Tell the user to label it as AI-generated wherever they publish it (Instagram has an "AI label" setting; the instagram skill sets it with post reel --ai-label), and to check the platform's current rules themselves. This skill never publishes.

What it needs

Run doctor first. It checks ffmpeg/ffprobe, the ComfyUI server (loopback only), the node classes and the seven model files (about 45 GB). When something is missing, read references/setup.md: it lists every file, its source and licence, and the one node pack to install. Ask before downloading; state size and source. The kept runs used a 32 GB GPU; less memory is untested.

The workflow

Read references/workflow.md once before the first remake: it has the decision rules, the measured timings and the failures behind each rule. In short:

1. Break the source down with the instagram-video-breakdown skill (shots, who appears, where). Decide per shot: keep as is, swap, or continue.
2. init the run folder with the source, ONE identity photo and the consent statement. Use the same photo for every shot; identity drifts when references change. init measures the cuts of the source to the frame; scenes <run> lists the shots as frame ranges. Check them against the breakdown (and fix them with scenes --set if a flash or a fast move was taken for a cut): every stretch is planned on these, not on approximate times.
3. Write the prompt from references/prompts.md: the same character card in every shot, plus that shot's framing, light and "keep the performance and everyone else".
4. swap each shot where the person appears, as --scene <n> (cut to cut; add --start to begin later, it still ends on the cut). One stretch per run, up to 10 s. A stretch that crosses a cut, or stops or starts within 8 frames of one, is refused: those frames would show the original person for an instant at the cut.
   • Close or medium shot, person alone or clearly the largest: --tier 768, default detection.
   • Other people in frame: add --point x,y on the person, in source-frame pixels.
   • Small or distant person: --crop w:h:x:y around them (9:16, about 2-3x their height), --tier 480, --point, --expand 32. Only the crop is regenerated.
   • Person only clearly visible at the END of the stretch (camera arriving, something in front): add --reverse. Tracking starts where the person is clear; the result is flipped back. Each run prints the first frame the model sees. Look at it: the point must be on the person.
5. place each shot: a cropped shot is blended back inside its own area (outside it the picture is the original); a full-frame shot is fitted to the frame. The placed clip has exactly as many frames as the stretch it replaces, repeating frames where the source repeats them; it reports cadence (frames that stand still while the source moves). Then WATCH placed.mp4 in motion.
6. approve what is good. Re-run swap with one thing changed for what is not (a new attempt; nothing is overwritten). references/troubleshooting.md maps each defect to a change.
7. continue when the user wants the shot to go on (the shot must be approved first): describe the movement in order, in the same tempo words as the shot; the last 25 frames of the shot are anchored automatically.
8. assemble the timeline from stretches of the source and placed shots, in order. Never close a gap between two shots with a few frames of the source: in a shot where the person is replaced those frames show the original person, and assemble refuses them. Remake the shot so it reaches the cut instead.
9. audio --rights "<basis>": by default the source's own sound. If the remake is longer than the source, the music must carry on, never loop: give a longer recording of the same track (--file) and let --match find where the source's sound sits in it. --rights is always required: ask the user on what basis the video may be published with that sound, and write down their answer, not yours. "Review copy only, not for publication" is a valid answer.
10. compare, then have the user watch final.mp4 with sound before anything is published.

Show the user frames and the comparison video at every review point; do not describe them.

Commands

Generation takes minutes (about 45 s for a 2 s crop at 480, 4-5 min for 5 s at 768 on a 32 GB GPU). If a command times out, the job keeps running: run the same command again with --resume <prompt-id> (printed, and in status), or --resume last when no job id was ever printed. Never submit the same shot twice to "retry". One command at a time per run folder: a second one is refused while the first is working.

Rules that came from failures

• One identity photo, one character card, every shot. Do not improve the prompt between shots.
• Never paste a generated face or body onto the original by hand; regenerate the shot.
• Do not hand-edit graphs; change a command option. The kept settings are in the code.
• Re-detecting a lost subject (--recover) is only for shots with one person: it can jump to someone else.
• Plan stretches on the measured cuts (--scene), not on breakdown times: those are off by a frame or two, and one frame of the original person at a cut is visible.
• 24 fps material carried at 30 fps repeats one frame in five. swap sends the model only the distinct frames and place repeats them where the source did. Sending all of them let the model's own 30->24 resampling drop real frames and keep repeats: the motion stalled two frames in five. Keep that default (--all-frames turns it off) and read cadence after place.
• Changing a shot makes what was built on it out of date: its approval, any continuation of it, the timeline and the final file are cleared. Redo them; do not reuse an old final.mp4.
• Report honestly what was not checked. Tempo, lip sync and likeness are judged by the user's eyes and ears, not by a number.

Additional resources

• references/workflow.md — decision rules per shot type, timings, what was tried and rejected.
• references/prompts.md — character card, shot card, swap and continuation prompt structure.
• references/setup.md — models (files, sources, sizes, licences), the node pack, ComfyUI.
• references/troubleshooting.md — each visible defect or error and the one change to try.
• references/manifest.md — the ig.remake.v1 record and how to resume from it.
• assets/workflows/h3_swap.api.json — the character-swap graph (MIT, see its NOTICE).