BudgetPixelBudgetPixelOpen the Video Workshop

Workflow

Give any face performance to a new character

Someone pulls faces for eight seconds. Then a grandma, a CEO and a pug pull the exact same faces, with the same timing. Nobody's face is pasted onto anybody: only the movement carries over, through a face mesh.

Make a face meshWorks on people, animals and characters
Input clip: “Mime face study 1” by Justin Morrison, CC BY, cropped and turned into a face mesh. Everyone else in the film is an AI character.

How it's made

  1. Step 1

    Pick a face clip

    Someone pulling big expressions, close up and looking at the camera, 5 to 10 seconds. Use a clip you filmed or have the rights to. No clip? A video model can perform one for you from a second-by-second list of faces.

  2. Step 2

    Turn it into a face mesh

    Open Depth Map in the Video Workshop, switch Face mesh on and keep Grayscale. The grey video keeps the head and every expression and drops who the person is. The mesh is free on top of the depth map.

  3. Step 3

    Pick the new face

    Write one of the free BudgetPixel influencers’ @handles in the prompt, @mention your own character, or describe anyone. It even works on a pug.

  4. Step 4

    Write the timeline

    List the seconds where each expression happens: “from 2.5 to 4 seconds, a huge round O”. Without it the model copies the faces smaller and out of order.

  5. Step 5

    Make it

    In Edit Video, pick Seedance 2.5, upload the face mesh as the video to edit and paste the prompt. Test at 720p, then make the final at 1080p. Each new face is its own job.

The tools in it

Depth Map with Face mesh

Makes the face mesh from your clip

Edit Video

Gives the performance its new face

Seedance 2.5

The video model that follows the mesh

AI influencers

Faces to cast with an @handle, free

Music 3.0

A song with your own lyrics for the cut

API

Two calls: the mesh, then the new face

Questions

What is face expression transfer?
Taking the facial performance from one video (the expressions, their timing, the head moves) and giving it to a different face. The result shows a new character doing the same performance.
Is it a face swap?
No. The person in your clip never appears in the result. The face mesh keeps only how the face moves, so the model builds the new face from scratch, which is also why the old face does not leak through.
Which clips work best?
One face filling most of the frame and looking at the camera, with big, exaggerated expressions, 5 to 10 seconds on a steady camera. Crop a wide clip square around the face first.
Does it work on animals and cartoon characters?
Yes. A mesh made from a human face worked on a pug: mouth open, tongue out, a grin at the end. Describe the character and the model maps the performance onto it.
Can I do it from code?
Yes, in two API calls: make the depth map with face_mesh on, then send its video to Seedance 2.5 as a reference video, with the new face in the prompt.

More workflows

Barry — IN MY HEAD vs REALITYIN MY HEAD vs REALITYSplit-screen comedy: the life your character imagines on top, the one they live below, cut on the beat.

Your clip, any face

Make the face mesh first: it takes about a minute and a half for eight seconds.

Make a face mesh