How to swap the people in any video with Higgsfield Genjutsu

Pick the clip first

Everything downstream inherits the source: its moves, its framing, its cuts. A weak original can’t be fixed later. What worked for me was a two-person performance on a plain set, both people facing the camera, only a couple of cuts.

• At least 5 seconds. The stated minimum is 4. Every one of my exactly-4-second clips processed for about three minutes and then failed. 5 and 6 worked every time.
• 30 seconds max. Anything longer gets trimmed to 30.
• You pay per second of source, so trim to the part you actually want.

# grab the source (any yt-dlp site works)
yt-dlp -f "bv*[vcodec^=avc1][height<=1080]+ba[ext=m4a]" \
  --merge-output-format mp4 -o source.mp4 "https://www.youtube.com/watch?v=x9yop0nYR9g"

# cut the section you want. 5-30 seconds. -ss is the start, -t the length
ffmpeg -ss 13.4 -i source.mp4 -t 20 -c:v libx264 -crf 16 \
  -c:a aac -movflags +faststart clip.mp4

One good photo per person

My first test used two photos of myself: one in ski goggles, one in sunglasses, small in the frame. It came back as two guys who looked nothing like me or like each other. The photos were the problem, not the model.

• Face big and clear, nothing covering it. No sunglasses, no hats, no hands near the face.
• Sharp, at least ~1000px tall. My 400px photos gave mushy faces.
• Full body if you can find one. A chest-up photo leaves the model guessing the build.
• The outfit in the photo is the outfit in the video. Ski jacket in, ski jacket out.

For public figures, Wikimedia Commons has sharp, properly licensed portraits. That’s where both of mine came from.

Say who plays whom, and lead with the fix

Assign each photo by position (“the man on the left becomes the man in image 1”) and describe what matters: hair, build, clothes. Two things I learned the hard way:

• Put the negatives first. The original performers wore sunglasses. “No glasses” at the end of the prompt got ignored. Moved to the first line, it worked.
• Describe the build in comparison. “Stocky” alone did little. “Noticeably heavier-set than the man on the left” did more.

Neither man wears glasses or sunglasses; both faces are fully uncovered.
The man on the left becomes the man in image 1: his face, no glasses, short dark
swept-back hair, clean-shaven, lean build, blue corduroy overshirt open over a white
ribbed tank top, thin silver chain necklace.
The man on the right becomes the man in image 2: his face, no glasses, short dark
hair, light stubble, stocky, broad-shouldered, muscular build with a thick neck and
full chest, noticeably heavier-set than the man on the left, charcoal blazer over a
black-and-white gingham shirt.
Keep the motion, hand gestures, timing, camera, orange set and microphone unchanged.

Test 5 seconds at 480p first

Price is per second of source video, rounded up: $0.318 at 480p, $0.681 at 720p. A 5-second test at 480p is $1.59. That buys you the only judgement that matters, whether the recast works at all, before you spend on the full clip.

Then run the whole thing, and wait

The full 20 seconds at 480p was $6.36. It took about 50 minutes: roughly 24 in the queue, the rest rendering. Start it and walk away. Download the result straight away, because outputs are only kept for about a week.

I stayed at 480p. Posted as a Reel with black bars, you can’t tell.

Post it

• Keep the original audio. The mouths are synced to it. Swap in a different sound and the lip-sync falls apart, which is the whole effect.
• Letterbox it. A landscape two-shot doesn’t crop to vertical without cutting someone in half. Center the 16:9 clip in a 9:16 frame and use the black bars for your text.
• Turn on the AI label. It’s a realistic video of real, recognizable people. Instagram and TikTok both require it.