Enjoying this issue?
Get tomorrow's AI & engineering digest in your inbox — hand-picked, summarized, and always spam-free.
TLDR
Higgsfield's Genjutsu is a video-to-video model that lets you edit specific elements in a clip—swap an object, person, or background—without regenerating the rest of the frame. The presenter demonstrates that a reroll costs 65 credits versus a full reshoot day, making it a practical tool for teams who need to patch a master instead of booking a crew again. The key limitation is lighting consistency when swapping backgrounds, but the ability to keep the original camera movement and timing intact is a significant workflow improvement.
Key points
Genjutsu has two modes: object swap (replace one element) and motion transfer (keep movement, rebuild everything else).
The presenter swapped a can for a bottle, a person for another person, and a colonnade for a beach boardwalk in the same source clip.
The demo cost 781 credits for two source clips and six finished swaps, with each swap costing 65 credits.
A lighting mismatch occurred when swapping backgrounds, requiring a reroll instead of a reshoot.
The tool is positioned as a way to edit video like patching software, avoiding the cost of regenerating entire clips.
Tools mentioned
Techniques
- object swap
- motion transfer
- reference image editing
Stop scrolling. Start reading smarter.
Receive the day's most important AI & engineering updates in one concise email. No spam.
Transcript (captions)
Watch the rug. Watch the fan behind her. Watch the plates laid out. Now watch the woman sitting there turn into a cat. Nothing else in that shot moved. Not the rug, not the fan, not the plates. That's
Genjutsu, the video-to-video model from Higgsfield, who are sponsoring this video. Genjutsu takes a clip you already have and edits inside it. One element out,
one element in. The rest of the frame stays where it was. Here's the same trick on a crowd. Green jackets become suits, and every person is still mid-step in formation.
And here's a face resiled from the jacket up, holding the same pose in the same terminal. So here's the plan. We'll build one clip on Higgsfield, change one thing inside it four separate times,
then read the credit meter. If you ship software, you know this shape. You don't rebuild the binary to bump one dependency. You patch it, then check what else moved. So when you patch a
video, what else moves? That is the whole video, and you only find out by trying it on footage you made yourself. The last video on this channel was about the cost of trying again. Not the price
on the pricing page, the price of the tries you throw away before one shot is good enough to keep. Roundups put the keeper rate at roughly one in four. So if a generation costs you a dollar, a
usable generation costs you closer to four, and the pricing page never mentions it. Every fix for that so far has been about making the next try better. Tighter prompts, lock seeds,
reference images, you're still rolling the dice just with better odds. Genjutsu is a different move. You stop asking for a better clip. You keep the clip you already like and change only the part
that's wrong. Which means the parts you liked can't drift because they're never regenerated. So let's make something to break. We generated a 9-second clip on Higgsfield first, so the source is ours
and you can see exactly what went in. That one cost 81 credits. A woman in an olive jacket walks a concrete colonnade at golden hour, lifts a dark can, and looks at it. The camera pulls back by
hand, so it wobbles. Remember that wobble. Then you open Genjutsu, and it hands you two modes. Object swap wants a reference video to edit. Motion transfer wants a reference video to extract
motion from. Same box, completely different job. Object swap is the find and replace one. You drop in the source clip, anything from 4 to 30 seconds, then up to 30
reference images of what you want instead, and you write one line describing the change. Ours read like this, "Replace the product in their hands with the one in the reference
image. Keep the person, their hands, the grip, the lighting, and the camera move exactly as they are." Notice what that prompt is doing. It isn't describing a scene. It names one thing to change,
then lists everything to leave alone, which is the same sentence you'd write in a bug report. And before you commit, the price is already sitting on the button, 65 credits with the list price
struck out beside it. You can read the bill before you spend it, which is rarer than it should be. Here's the source on the left and the result on the right. The can is a glass bottle now. Her hands
haven't moved. The wobble is the same wobble, and the light down that colonnade is untouched. Now the harder one. Same clip, same jacket, same can, except this time the reference image is
a different human being. Watch the jacket. It's the same jacket on a different body, folding the same way as he takes the same step. The can sits at the same angle at the same moment in the
walk. That's the beat worth sitting on. A person is the hardest thing to swap and the easiest thing to get wrong. So, how much of a shot are you willing to hand over before you stop calling it
yours? Third swap, same source, and this time the reference is a place. The colonnade becomes a beach boardwalk at sunset and she keeps walking straight through it. She's still holding the same
can, still turning her wrist on the same second. The world behind her is somewhere else entirely and here's the one thing to watch. Look at the left of that frame. The sun sits low on the
left. The rail catches a hard edge off it and the subject is still lit from the right, the way the colonnade lit her. That mismatch is a reroll, not a reshoot. You swap the location reference
for one lit from the same side or you simply generate again and you're out 65 credits instead of a shoot day, which brings us to the second mode where motion transfer keeps the move and
rebuilds everything else so the take survives and the world around it does not. The prompt for that one is worth reading because the references are addressable. You point at image one for
the person, image two for the product, image three for the location and the model knows which is which. Here's the result beside the source. New person, new bottle, new beach, same walk, same
timing, same handheld drift down to the frame where the wrist turns. A second clip got the same treatment. A man at a kitchen table with a mug. Two more versions came out of one take. That's
the practical reason a team would care. You shoot once, then a product changes or a market wants a different face and you patch the master instead of booking the crew again. So, the bill. We opened
this session with 3,000 credits and closed it with 2,219. That's 781 credits for two source clips, six finished swaps and every reroll in between. Higgsfield keeps a promo
running on that page and the shelf next to it keeps moving, too. The plan panel lists Nano Banana Pro, Nano Banana 2 and Kling 3 all on 7-day unlimited. The offer attached to this video is the one
to use. Higgsfield is giving you unlimited generations for 7 days at 70% off, and the link is in the description below. Here's where I land. If you make video for a living, Genjutsu is the
first tool this year that treats a clip as something you edit rather than something you reorder, and it's worth the seat on that alone. It isn't magic, and that lighting note is real, but a
re-roll costs 65 credits, and a re-shoot costs a day, which raises the thing I keep chewing on. Once one element of a shot is editable, and then all of them are, what exactly did you shoot? That's
the take. Higgsfield sponsored this one, and the 70% off link is in the description below. Go break a clip on purpose.