A viral dance clip works because of its timing: the beat drop, the head turn on the snare, the camera cut on the last step. When we started building a character swap for ClipSpeed, the goal was simple to state and hard to hit: put your character in place of the person, and change nothing else.
This is what we learned getting there — including two approaches that looked right and weren't.
A finished swap — watch the video and see the template.
Lesson 1: "Generate the video again with my character" is the wrong model
The obvious approach is to treat a swap as video generation: give a model the original clip as a reference and ask for the same scene with a new person. The output often looks great frame by frame — and it's wrong. A model that regenerates the video re-imagines the motion: the dance gets smoother and more generic, hands land in different places, the mouth no longer matches the audio. You get a nice video that is no longer the trend.
The fix is a different class of model: one that tracks the original performance (body pose, face, expression, mouth) and re-renders only the person, while the real pixels around them stay. That's the difference between "a video inspired by the clip" and "the clip, with a different person in it."
Lesson 2: The character reference matters as much as the clip
Swap quality depends heavily on the character image. A face-only photo leaves the model guessing about the body, outfit and proportions — so the person in the video gets an improvised body.
So Morphaze checks: if the photo shows only a face, it first draws the full character, head to toe (a cheap image step), and uses that as the reference. Saved AI Influencer characters plug in directly, which is how one character can show up in a new trending format every day and still be recognisably the same.
Lesson 3: Long clips have to be made in pieces — and the joins are the whole problem
Swap models take a limited length of reference video per pass. Plenty of viral clips run 16–33 seconds, and trimming them to fit cuts off the payoff. So longer templates are rendered in pieces and joined. The naive versions of that fail in instructive ways:
- Cut where the source cuts. We plan piece boundaries at the clip's own camera cuts, so most joins happen where the viewer already expects a change of picture.
- Never trust nominal timestamps. A render keeps its length but can shift a camera cut by a frame or two. Each piece is matched back to the original frame by frame, and the switch happens on each render's own cut frame.
- Inside a single shot, simple transitions fail. A crossfade shows a "double" (two renders never frame a pan identically), and a hard one-frame switch makes faces and background details visibly pop. What worked: align the outgoing piece to the incoming piece's framing and colour, then run a short optical-flow morph across the best-matching pair of frames, then hand over to the next piece untouched.
- Measure, don't eyeball. A join is accepted only if its frame-to-frame jitter is no worse than the renders' own jitter, checked alongside frame strips around the seam.
The original audio is laid back over the joined result, so the swap lands on the same beat it came with.
Lesson 4: Some things the model just won't guess
- Multiple people need a plan per shot. Templates can hold two or three people, each replaced in their own place — but "replace everyone" as one instruction is how you get one person left unchanged.
- Empty scenery should stay original. Where the person leaves frame, a swap model may paste the character into the landscape, so those stretches use the original footage.
What it looks like to use
- Pick a template — a trending clip already prepared for a swap — or upload your own MP4 of 4–15 seconds that you have the rights to use (one main person, body in frame).
- Add a character — a saved AI Influencer or any photo.
- Generate — the character replaces the person and does exactly what they did, in 720p or 1080p HD, with the original sound.
It also runs from Claude or ChatGPT through ClipSpeed's MCP server (morphaze_swap), with the price shown before anything runs.
Cost: Morphaze is billed on clip length — about 48 credits for a 10-second clip at 720p, 96 in 1080p HD, from the same credit wallet as clipping (Starter $19 for 300 credits, Pro $39 for 600, Ultra $99 for 1,500). Morphaze unlocks when a plan starts.
Use it responsibly: only upload clips you made or have permission to use, only use characters you created or have the rights to, never use it to impersonate anyone, and label AI video where the platform asks you to.
More examples and the template list: clipspeed.ai/morphaze.
Questions about the joining pipeline are very welcome in the comments — it was the most fun part of the build.
I drafted this post with help from Claude; the facts, numbers and examples come from ClipSpeed itself.
Top comments (0)