The task
The traders’ club Full Hamster has a cartoon hamster from its stickers: flat, light brown, cream chest, big cheeks, two front teeth. The idea: take a clip with a large crowd of extras and replace everyone with crypto animals. Bears in emerald-green suits with red ties, bulls in bright red with green ties, all in dark glasses, and in the centre, instead of the man in black, our hamster. The donor: 27 seconds, extras in white suits on the steps of a Japanese temple, eight cuts. At the end, the caption ALTSEASON.
The same day, a second video with the same technique: a photorealistic bear and bull in suits rap over a donor video of two people on an orange background, with hand-drawn candles and arrows. Later the recipe went into gifts: a birthday girl in a crowd of bodybuilders, and a close friend in place of the dancer, the skydiver and the woman with the bottle in three donor videos.
Why, and what the idea was
Motion transfer gives what generation from scratch does not: exact choreography, head turns, lip sync and timing from the donor, with our own characters on top. Higgsfield has two different Motion Controls, and they are not equal: the old one on Kling 3.0 takes no prompt and holds one character; Genjutsu takes a prompt and up to 49 characters. So Genjutsu, always. The output length equals the donor length, there is no duration parameter, and there is no 10-second limit like in Kling 2.6.
Characters are made in advance in Nano Banana Pro and fed in as images, one figure on a neutral background each. The hamster stayed flat 2D next to photorealistic animals and never drifted once in 27 seconds: the model can hold two styles in one frame.
How it went
| When | What happened |
|---|---|
| Run 1 | Bear and bull in one picture. The output was almost all bears, suits handed out at random: the colour was tied to nothing |
| Run 2 | Two separate references, the prompt said “exactly 50 percent” and “count them as you place them”. The bulls came back, but at 10 to 20 percent, suits still mixed up, the temple still there |
| Run 3 | A bias, “bulls must clearly outnumber bears, at least two in three with horns”, a no-overlap rule as the first item, direct bans “no temple, no shrine buildings, no lanterns, no roofs”, a Latin caption spelled letter by letter. Accepted as final |
| In parallel | “A bear and a bull rap”: two tries at 87 credits each, the final accepted. The hand-drawn graphics came out better than expected. The Cyrillic caption came out as gibberish, the Latin FULL HAMSTER without errors |
| Three days later | The same scheme for a birthday: a crowd of bodybuilders, the birthday girl from a photo in the centre. Bare shoulders, straps and a small floral pattern drifted in the first seconds, the caption came out as MARGU. Second run: a plain red dress with sleeves, the caption split into two lines |
| Three weeks later | A gift for a close friend’s 54th birthday: three donors (a dance of 32 s, a skydive of 17 s, a bottle of 9 s), a character sheet with FACE LOCK, a separate prompt per video. The bottle donor has the number 35 baked in, so that prompt carries the strictest ban on text, and 54 was overlaid in the edit with the same slide-in as the original |
Tools and method
The call: generate_video, model hf_mult_motion_control, images with the role image and exactly one video. The platform offers a preset such as “DROWN IN MUSIC” and will not start without a refusal: we decline the preset, otherwise it replaces the idea. The price depends on donor length and resolution: 27 seconds is 81 credits at 480p, 190 at 720p, 299 at 1080p, roughly 3, 7 and 11 credits per second. A Nano Banana Pro image is 2 credits. A run takes 20 to 45 minutes; that is not a hang.
The prompt is one block in English, with the bans woven into the text. Order matters, because the model holds the beginning and the end best: first the no-overlap rule, then crowd proportions through bias, a hard tie of clothing to character, a description of each with its reference number, the background with bans on what must go, graphics and the final caption, and at the tail photorealism, no humans, no watermarks.
For a real person in the reference the clothing is closed and plain, because bare skin and small patterns drift. For the face, a FACE LOCK block: shape, cheekbones, eyelids, wrinkles, age, and direct bans “do not narrow the face, do not make him younger, do not blend two faces into one”. Bans like “no grey hair” break the face, so hair shade is changed only through the HAIR line.
Result
ALTSEASON: 27 seconds, 854x480, 24 fps, shown here without sound because the donor carries someone else’s track. What works: the caption has no errors, the temple is gone and the sky has opened, bulls are about half, the hamster cuts off nobody’s head, the cartoon look holds for the whole video, the suits are mostly right. What remains and was accepted as is: occasional suit errors, a bear in red at second three, no white hand-drawn graphics on the background, the model only delivered the caption, and distant figures are blurred in wide shots.
“A bear and a bull rap”: 29 seconds, 752x560, the mute version. The motion transferred exactly, the hand-drawn graphics appear and vanish, the FULL HAMSTER caption is clean. Two blockers before publishing: the format is not vertical and the original carries the donor’s audio, so for TikTok and Reels it goes up without sound with a track from the platform’s library.
Credits: 255 for ALTSEASON, three runs at 81 plus 12 for images, and 178 for the rap video. The animal references are ready and uploaded, so the next video on the same scheme is 81 credits and about an hour of work. What we do not know: the result of the second birthday run was not recorded, and the gift videos for the friend were launched from our instructions on the client’s side.
Mistakes and what we took from them
