HOW BULIN 47 AI WORKS
How Bulin 47 AI Works, Part 1: Why We Built on Recast, Not Prompts
This is part 1 of a 3-part series on the technology behind Bulin 47 AI. Part 1 explains the core idea and the model. Part 2 follows a video from upload to download. Part 3 covers quality, limits and what the benchmarks do and do not tell you.
When the Bulin 47 freestyle exploded in October 2026, most people who tried to recreate it with AI hit the same wall. They opened a text-to-video tool, typed something like "a rapper freestyling on a Madrid street with a speaker and a crowd," and got a video of... a rapper freestyling on a street. Pleasant, generic, and nothing like the clip everyone was talking about.
That gap, between a freestyle and the freestyle, is the reason Bulin 47 AI exists, and it shapes every technical decision we made.
The problem with prompting a viral moment
A viral clip is not a concept; it is a specific performance. The beat, the shoulder rhythm, the way the partner holds the speaker, the framing of the street, the sound of the original take: none of that survives a text description. You can describe "a rapper dancing," but you cannot describe that dance.
Image-to-video has the opposite weakness. Give a model one photo and it will invent motion, which means the movement is the model's idea, not the original performer's.
So the question we asked was different: instead of asking a model to imagine the freestyle, what if we kept the real freestyle and only changed who is in it?
What "recast" means
That is exactly what a recast model does. Recasting is a video-to-video task. According to fal's documentation for H3 Max Recast, it replaces the main people in a source video with people from reference photos, tracking each person across shots, while preserving the source motion, camera, cuts and audio.
Read that list again, because it is the whole product in one sentence:
- Source motion is preserved, so the choreography stays the original performer's.
- Camera and cuts are preserved, so the framing and the street-level feel stay intact.
- Audio is preserved, so the video keeps the original sound rather than a regenerated approximation.
- People come from reference photos, so the identity on screen is the one you upload.
The Bulin 47 trend is, by definition, a character-replacement remix: the lead performer is swapped, while the partner, the crowd, the lighting and the camera movement stay as they were. A recast model is the technical version of that same idea.
The division of labor: clip carries the performance, photo carries the identity
It helps to think of two inputs doing two different jobs.
| Input | What it controls |
|---|---|
| Prepared source clip | Choreography, timing, camera, setting, other people, audio |
| Your portrait | Face, hair, clothing, overall appearance of the replaced performer |
Because the performance comes from a real recording, there is nothing for the user to describe and nothing for the model to improvise about the dance. That is why Bulin 47 AI has no prompt box. The part prompts are good at (describing a scene) is already solved by the source clip, and the part prompts are bad at (reproducing an exact performance) is handled by the recording itself.
The engine: H3 Max
Bulin 47 AI is powered by H3 Max Recast, served through fal. A quick note on what H3 Max is, because the name causes some confusion.
H3 Max is a post-trained version of the open-weights MiniMax H3 video model, developed by fal Research and optimized for speed by fal's inference team. fal reports that it was post-trained for stronger prompt adherence and visual quality and co-optimized with a custom inference stack. Its headline claim is generation of a five-second video in roughly three seconds.
On quality, fal reports that H3 Max ranked first in its human-preference evaluations for overall quality, prompt understanding and aesthetics against twelve other video models, including Seedance 2.5, Kling 3 and Veo 3.1. fal also reports first-place results in image-to-video categories on independent leaderboards from Artificial Analysis and Design Arena.
We will be careful about how we use those numbers, and Part 3 explains why. For now, the important point is simpler: H3 Max Recast is a model built specifically for the "keep the video, change the people" job, on top of a base model that is currently among the most talked-about in video generation.
Why we chose a single prepared template
Most video tools let you upload any clip. We deliberately did not. Bulin 47 AI uses one prepared source clip, cut into 6-, 12- and 19-second versions, rendered in vertical format.
That choice has real engineering benefits:
- Predictable quality. We know exactly what the source footage looks like: where the performer stands, how fast the shoulders move, where the speaker is. The only variable is your photo.
- Fewer failure modes. An open "upload any video" tool must cope with unknown lighting, unknown camera moves and unknown numbers of people. A fixed template removes most of those unknowns.
- Simpler product. A user who only has to pick a photo and a length can finish in under a minute of effort.
- Rights-aware by design. A prepared template keeps the experience tied to the one trend it was built for, instead of becoming a general tool for remixing anyone's footage.
The trade-off is just as clear: if you want a different clip, this is not the tool. It is a specialist, not a studio.
What comes next
In Part 2 we follow a single video from the moment you upload a portrait to the moment you download the MP4, and explain why generation is asynchronous, what the queue does and what happens when something goes wrong.
Want to see the result of all this? Try Bulin 47 AI →
Bulin 47 AI is an independent fan-made project and is not affiliated with or endorsed by Bulin 47 or Kevin Wai.
---
Sources and further reading
- fal, "Introducing H3 Max by fal": https://blog.fal.ai/introducing-h3-max-by-fal/
- fal model page, H3 Max Recast: https://fal.ai/models/minimax/h3-max/recast