Video editing direction · 06 Nov to 12 Nov
Levon Khachatryan · follow-up to the 05.11 deck
H3 video editing: face swap and style transfer
The two use cases we focused on this week: swapping the people in trending videos, and turning a video into another style.
1 · Face swap on trending videos
3 viral clips, the whole 15 s in one pass; H3 arms and the new Motion Transfer LoRA
2 · Video style transfer
114 cases (real, animated and art styles), 9 methods
Where we are
We focus on two video-editing use cases
Based on the previous week's investigation (h3-video-editing.pages.dev) and the discussion in our meeting on 5 November, we decided to focus on two use cases for video editing for now:
Use case 1
Face swap for trending videos
Replace the people in a viral clip with new people from a picture.
Use case 2
Video style transfer
Turn a video into another look (anime, clay, sketch, live action and more) while keeping its people, motion and timing.
Part 1
Face swap on trending videos
Three viral clips, the first 15 s of each, new people from GPT Image 2 pictures, every method in one pass.
Setup
Three trending videos, five H3-based methods
Dance trend
3 women in a row become 3 grandmothers; 3 pictures, vertical clip.
Love of the Dead
A movie-style couple, each in 2 looks (house and field); 4 pictures.
H3 Ref2VA
base H3, FastH3, Pika H3 Flex: source as <Video 1> plus the pictures, who-is-who in a shot-by-shot prompt.
New: Motion Transfer LoRA
Community LoRA for H3 in two versions, TARA and Guide Latents; commercial use allowed (credit the author above $1M revenue).
All at 768p, 15 s (362 frames), one pass with all pictures. We also tried two community character-swap and head-swap LoRAs: they take one person per run, so multi-person clips need chained passes; that did not work well, so we leave them out for now.
Swap · dance trend
Three women, three grandmothers
Source | base H3 | FastH3 | Pika | Motion Transfer TARA | Guide; the 3 pictures below with who and when.
What to see
Both Motion Transfer versions swap each woman at her moment and follow the dance. base H3 and FastH3 show the grandmothers, but not in the source’s order and timing; Pika keeps the originals.
Swap · dance trend
Motion Transfer TARA
Condition pictures | result, one pass with all three pictures.
What to see
The swap holds for the whole 15 s, each grandmother replaces the right woman, and the parking lot and camera of the source stay.
Swap · dance trend
Motion Transfer Guide Latents
Same pictures, the other Motion Transfer version, one pass.
What to see
The same three swaps and the same dance; the setting is redrawn (a petrol station instead of the open parking lot). Also usable.
Swap · dance trend · 2nd example
Three grandfathers: TARA
New GPT Image 2 pictures, same clip and roles, one pass.
What to see
Each grandfather appears in his own shot and keeps his look: the bearded dancer in the tracksuit, the man with glasses taking the selfie, the man in the flat cap. No women remain.
Swap · dance trend · 2nd example
Three grandfathers: Guide Latents
Same pictures, the other Motion Transfer version.
What to see
The same three swaps with identities kept; as before, the setting is redrawn as a petrol station.
Swap · movie scene
Love of the Dead
A couple, each in two looks: the man in the house and in the field, the woman as a zombie and in the field.
What to see
The hardest clip: many short shots and two looks per person. Motion Transfer follows the source shots; base H3 and FastH3 invent their own shots.
Swap · movie scene
Motion Transfer Guide Latents
Result left, the four pictures right; one pass.
What to see
Both new people appear in their shots and the cuts and poses follow the source.
Swap · movie scene
Motion Transfer TARA
Same pictures, the other Motion Transfer version, one pass.
What to see
TARA keeps the source rooms and light closely; check per shot whether each new person replaces the original actor.
Swap · movie scene · 2nd example
A Black couple: Guide Latents
Distinctive faces, so identity is easy to judge; each person in two looks.
What to see
Correct in every shot: the man keeps his face, twists and beard; the woman keeps her afro, also in the zombie look.
Swap · movie scene · 2nd example
A Black couple: TARA
Same pictures, the other Motion Transfer version.
What to see
Identities hold as with Guide Latents, except one short shot: in the second zombie close-up (about 9 s) the original actress leaks through.
Swap · limitation
A dog in a human role does not work
A Chinese man and a husky (normal and zombie look) in place of the couple; TARA shown.
What to see
The man is right in every shot. The dog appears only in a few shots; elsewhere the original woman stays or a second man appears. Guide Latents is worse. A dog works when the motion is person-like (the upright dog on Migos), not in human poses and actions.
Swap · music challenge
Migos lobby challenge
Source | base H3 | FastH3 | Pika | Motion Transfer TARA | Guide Latents: two performers at a microphone become a new couple (woman left, man right).
What to see
All five methods swap both people, but the two Motion Transfer versions, TARA and Guide Latents, follow the source best; base H3 loses the timing in places. From here on we show only these two.
Swap · music challenge
Motion Transfer TARA
Result left, the two pictures right; one pass.
What to see
The new couple performs every move of the source, in sync with the song, in the same room as the source.
Swap · music challenge
Motion Transfer Guide Latents
Same pictures, the other Motion Transfer version, one pass.
What to see
The same swap and the same moves in sync with the song; the set is redrawn with a red curtain instead of the source room. Also usable.
Swap · music challenge
TARA on 3 more couples
Same 15 s Migos clip, TARA, one pass; the two pictures in each corner.
What to see
Every couple performs the source’s moves for the whole clip, even a man and a dog standing upright. The result holds beyond one lucky pair.
Swap · music challenge
Guide Latents on the same 3 couples
Same clip and pictures, the other Motion Transfer version, one pass.
What to see
The swaps and the motion are as good as with TARA; the difference is the set, which Guide Latents replaces with a curtain. Both versions are usable.
Face swap · what we see
Motion Transfer is the first method that swaps from start to end
Motion Transfer
Both versions swap the new people in for the whole 15 s and follow the source motion on the dance and Migos, plus a second dance and Love cast and 3 more Migos couples. Limitation: a dog can replace a person only when the motion is person-like. Both versions are usable: TARA keeps the source set, Guide Latents often replaces the background.
base H3 · FastH3 · Pika
Swap well on a simple two-person shot (Migos), but on the dance and Love they lose the source’s order, timing or shots.
Single-person LoRAs
The community character-swap and head-swap LoRAs need one pass per person; chained passes did not work, so we leave them out for now.
Three clips, one sample each: a first look, not a ranking.
Part 2
Video style transfer
Real to style, style to real and style to style: 114 cases, H3 against community H3 LoRAs and open editing models.
5 s · to ↓ / from →
Real
3D animation
Anime
2D cartoon
Claymation
Comic
Video game
Painted anim.
Pixel art
Puppets
Toy bricks
2D cartoon
Cel-shaded game
Low-poly
Film noir
Pop art
Marble statue
3D animation to real · 5 s · 4 examples
Anime to real · 5 s · 8 examples
2D cartoon to real · 5 s · 3 examples
Claymation to real · 5 s · 4 examples
Comic to real · 5 s · 3 examples
Video game to real · 5 s · 3 examples
Painted animation to real · 5 s · 3 examples
Pixel art to real · 5 s · 3 examples
Puppets to real · 5 s · 3 examples
Real to 3D animation · 5 s · 2 examples
Anime to 3D animation · 5 s · 1 example
2D cartoon to 3D animation · 5 s · 2 examples
Comic to 3D animation · 5 s · 2 examples
Real to anime · 5 s · 4 examples
3D animation to anime · 5 s · 1 example
Claymation to anime · 5 s · 2 examples
Video game to anime · 5 s · 2 examples
Real to toy bricks · 5 s · 2 examples
Real to 2D cartoon · 5 s · 2 examples
Real to cel-shaded game · 5 s · 2 examples
Real to claymation · 5 s · 2 examples
Anime to claymation · 5 s · 2 examples
Pixel art to claymation · 5 s · 2 examples
Real to comic · 5 s · 2 examples
3D animation to comic · 5 s · 2 examples
Painted animation to comic · 5 s · 2 examples
Real to ink wash · 5 s · 2 examples
3D animation to ink wash · 5 s · 2 examples
Puppets to ink wash · 5 s · 2 examples
Real to low-poly · 5 s · 2 examples
Real to film noir · 5 s · 2 examples
Real to oil painting · 5 s · 2 examples
Anime to oil painting · 5 s · 2 examples
Claymation to oil painting · 5 s · 1 example
Video game to oil painting · 5 s · 2 examples
Real to pixel art · 5 s · 2 examples
Anime to pixel art · 5 s · 2 examples
Claymation to pixel art · 5 s · 2 examples
Real to pop art · 5 s · 2 examples
Real to pencil sketch · 5 s · 2 examples
Painted animation to pencil sketch · 5 s · 2 examples
Puppets to pencil sketch · 5 s · 2 examples
Real to marble statue · 5 s · 2 examples
Real to ukiyo-e · 5 s · 2 examples
2D cartoon to ukiyo-e · 5 s · 2 examples
Pixel art to ukiyo-e · 5 s · 2 examples
Real to watercolour · 5 s · 2 examples
Anime to watercolour · 5 s · 3 examples
Comic to watercolour · 5 s · 2 examples
15 s · to ↓ / from →
Real
3D animation
Anime
2D cartoon
Claymation
Comic
Video game
Painted anim.
Pixel art
Puppets
Toy bricks
2D cartoon
Cel-shaded game
Low-poly
Film noir
Pop art
Marble statue