In early October, a spontaneous street freestyle by Bulin 47 and Kevin Wai in Madrid went viral. Media reports put it in the tens of millions of views, and TikTok quickly filled with AI remixes that swap the lead performer for someone else. I tried a bunch of them and ran into an interesting technical problem, so I built a small site around it:
https://bulin-47-ai.org/
What it does
Upload one portrait, pick a 6, 12 or 19-second cut, and a few minutes later you get a vertical video of yourself in that freestyle, with the original motion, camera, setting and audio intact.
Why not text-to-video or image-to-video?
I tried the prompt route first and it doesn’t work for this:
Text-to-video gives you a street freestyle, never that freestyle. The rhythm, the shoulder movement, where the partner holds the speaker, the whole scene: none of it survives a text description.
Image-to-video invents the motion, so the movement is the model’s idea, not the original performer’s.
So I flipped the question. Instead of asking a model to imagine the performance, keep the real recording and only change who is in it.
The technical approach
It calls minimax/h3-max/recast on fal, a video-to-video model. According to fal’s docs, it replaces the main people in a source video with people from reference photos, tracks each person across shots, and preserves the source motion, camera, cuts and audio.
H3 Max is fal’s post-trained, inference-optimized version of the open-weights MiniMax H3 model.
There is exactly one prepared source clip (cut into three lengths, vertical). No prompt box. The user’s only input is a photo and a duration.
Generation is asynchronous: jobs go into a queue, results appear in a “My Videos” page, and credits are returned if a generation fails.
The fixed template is a deliberate trade-off. It gives predictable quality, fewer failure modes, predictable cost and easy testing. The price is that it only does this one trend.
Honest limitations
Overall I’m happy with the results: the audio, camera and choreography are the real ones, not regenerated approximations.
It isn’t perfect. Hands, fast movement, clothing edges and moments of occlusion vary between runs, and lip-sync isn’t guaranteed to be exact.
Photo quality matters a lot. Bright, front-facing, one person works best. Group shots, heavy filters, big sunglasses and tiny crops make results noticeably worse.
On “is it the best?”: fal reports H3 Max ranking first in its own human-preference evals and in some image-to-video categories on independent leaderboards. But that’s general video generation, and I’m not aware of a public benchmark for character replacement specifically. So I won’t claim “best”. Try it and compare.
Disclaimer
This is an independent fan project, not affiliated with or endorsed by Bulin 47 or Kevin Wai. Please only upload images you have the right to use, no minors, and label the result as AI-generated where platforms require it.
Feedback I’d love
Is a one-template, one-trend tool the right shape, or should it grow into multiple templates?
Video generation isn’t cheap. Beyond credits, what’s a sensible free tier / anti-abuse approach?
Anyone with experience on SEO and cold start for trend-driven sites?
Happy to answer questions in the comments.
https://bulin-47-ai.org/
What it does
Upload one portrait, pick a 6, 12 or 19-second cut, and a few minutes later you get a vertical video of yourself in that freestyle, with the original motion, camera, setting and audio intact.
Why not text-to-video or image-to-video?
I tried the prompt route first and it doesn’t work for this:
Text-to-video gives you a street freestyle, never that freestyle. The rhythm, the shoulder movement, where the partner holds the speaker, the whole scene: none of it survives a text description.
Image-to-video invents the motion, so the movement is the model’s idea, not the original performer’s.
So I flipped the question. Instead of asking a model to imagine the performance, keep the real recording and only change who is in it.
The technical approach
It calls minimax/h3-max/recast on fal, a video-to-video model. According to fal’s docs, it replaces the main people in a source video with people from reference photos, tracks each person across shots, and preserves the source motion, camera, cuts and audio.
H3 Max is fal’s post-trained, inference-optimized version of the open-weights MiniMax H3 model.
There is exactly one prepared source clip (cut into three lengths, vertical). No prompt box. The user’s only input is a photo and a duration.
Generation is asynchronous: jobs go into a queue, results appear in a “My Videos” page, and credits are returned if a generation fails.
The fixed template is a deliberate trade-off. It gives predictable quality, fewer failure modes, predictable cost and easy testing. The price is that it only does this one trend.
Honest limitations
Overall I’m happy with the results: the audio, camera and choreography are the real ones, not regenerated approximations.
It isn’t perfect. Hands, fast movement, clothing edges and moments of occlusion vary between runs, and lip-sync isn’t guaranteed to be exact.
Photo quality matters a lot. Bright, front-facing, one person works best. Group shots, heavy filters, big sunglasses and tiny crops make results noticeably worse.
On “is it the best?”: fal reports H3 Max ranking first in its own human-preference evals and in some image-to-video categories on independent leaderboards. But that’s general video generation, and I’m not aware of a public benchmark for character replacement specifically. So I won’t claim “best”. Try it and compare.
Disclaimer
This is an independent fan project, not affiliated with or endorsed by Bulin 47 or Kevin Wai. Please only upload images you have the right to use, no minors, and label the result as AI-generated where platforms require it.
Feedback I’d love
Is a one-template, one-trend tool the right shape, or should it grow into multiple templates?
Video generation isn’t cheap. Beyond credits, what’s a sensible free tier / anti-abuse approach?
Anyone with experience on SEO and cold start for trend-driven sites?
Happy to answer questions in the comments.
Select Language