See what’s worth creating.
Real video experiments, with the inputs, results and tradeoffs left in.
From drawn action to a new performer.
The broad action carries through. Faces and proportions still change.
Drawn action → photographic character
Compare Drawn action → photographic character
Inspect movement closely
Both videos stay on the same timeline at every speed; this does not retime the character’s moves. Frame buttons seek by 41.7 ms on the output grid; browser decoding and a different source frame rate can affect the displayed frame.
Loading the comparison videos… Open the result file
Silent previews. One timeline keeps the same moment aligned.
What to look for Guard, crouch, palm strike and return carry through. Face identity, muscle proportions and the cartoon energy effect do not.
Find your next experiment.
Choose a question. Watch the evidence. Keep what is useful.
01Watch comparison Smaller motion inputs. Shorter waits.
Two complete comparisons cut recorded run charges by 26–38%. Faces and hands still need work.
Tested October 14, 2026 · Saved results
02Watch comparison Keep the worker. Skip another startup.
A live handoff worked. The shared eight-second generated section was byte-identical; waiting in line was still real.
Tested October 28–29, 2026 · Saved results
03Watch comparison More seconds. The same scene?
An eight-second source became 16- and 24-second videos. The original audio survived; new footage was silent.
Tested October 21–28, 2026 · Saved results
04Watch comparison A dance can cross styles. Details don’t always follow.
Two useful motion-transfer examples. Neither is an exact identity or whole-scene conversion.
Tested October 15, 2026 · Saved results
05Watch comparison More references aren’t a face-swap guarantee.
H1 was faster. HM4 avoided one copied backdrop, but neither reliably reproduced the celebrity’s face.
Tested October 19, 2026 · Saved results
06Watch comparison A bigger encoder didn’t fix the face.
Two matched comparisons kept the movement. Neither showed a convincing identity improvement.
Tested October 29, 2026 · Saved resultsEarlier demonstrations
Historical examples with different models and workflows. They are not controls for the studies above.