If you have been on TikTok or Reels this October, you have seen the Bulin 47 clip: the Dominican artist Bulin 47 improvising a dembow freestyle on a street in Spain while Kevin Wai holds a portable speaker, and a crowd films it under red lights. The duo later released the track "Dame Hombro" (October 7, 2026), and the clip turned into an AI trend: people replace Bulin 47 with themselves, a friend, or a pet, and keep everything else. We built a small one-photo tool for this trend and spent the last two weeks looking at real user results. This post is not a feature list. It is a set of practical notes on what makes a Bulin 47 AI video look like the person in the photo, because that turned out to be the hardest part, and the place where most results go wrong.
Video editing, not image-to-video There are two ways to make this kind of clip:
Image-to-video: give a model a photo and a prompt like "a man freestyling outside a club with a speaker". You get a new scene. It may be fun, but it is not the clip: the camera, the crowd and the timing are different.
Video editing (subject replacement): give the model the original clip as a reference plus a photo, and ask it to replace one person. The camera movement, crowd, lighting and rhythm stay the same. This is what people actually mean by "the Bulin 47 AI video".
We use Seedance 2.5 reference-to-video in its editing mode: the reference clip is passed as [Video1], the user's photo as [Image1], the aspect ratio is set to follow the reference, and the output length follows the reference clip (we offer 15 and 30 seconds).
Lesson 1: strip the audio from the reference clip
Our first attempts with the original audio left in the reference were refused. The fix was simple: send a silent reference clip and generate without audio. If you want the original sound in the final video, add the audio track back after generation by remuxing it into the MP4 (no re-encoding needed). We deliver two files: one with the original sound and one silent copy for people who prefer to pick the sound inside TikTok.
Lesson 2: tell the model exactly who to replace
The original clip has several people in frame. If the prompt just says "replace the man", the model sometimes picks the wrong person. Describe the target by position and clothing, not by name: the performer in the open pale-pink shirt, standing behind the person holding the portable speaker.
Lesson 3: sunglasses cause closed eyes
Bulin 47 wears dark sunglasses in the clip. When the model removes them to show the new face, it often has no idea what the eyes should look like, and early results had closed or half-closed eyes. One line in the prompt fixed most cases:
His eyes are clearly open and natural, with normal blinking, looking at the camera and the crowd. He wears sunglasses only if he wears them in the photo.
Lesson 4: full-body photos are the main reason for "it doesn't look like me"
This was the big one. Our most common complaint was a face that looked rounder or "averaged". When we compared the uploads with the results, the pattern was clear: people upload full-body photos, where the face is maybe 5–10% of the image width. The model gets very few pixels of face, so it fills the gaps with the original performer's features. Expressions matter too: the original performer laughs a lot, and a laughing face is wider.
Three changes made a visible difference:
Ask for the right photo. A clear, high-resolution, chest-up photo, facing the camera, in even light, with no sunglasses and no other people. We show four small examples next to the upload box (one good, three common mistakes).
Auto-crop in the browser. If the face is small in the frame, we detect it with a lightweight face detector (about 0.5 MB, running locally) and crop the photo to chest-up before sending it. The user can still choose the original.
Send a face close-up as a second reference. From the same photo we cut a square close-up of the face (about 1.8× the face width, including hair, ears, beard and chin, resized to 1024×1024) and pass it as [Image2]. The prompt then says that [Image1] and [Image2] show the same person and that the face must match [Image2]. In our side-by-side tests this gave a much closer likeness than a single photo.
Lesson 5: describe the face positively
Our first likeness prompt was full of "do not": do not blend the original performer, do not make the face rounder. Results barely changed. Rewriting it as a positive description worked better:
His face matches [Image2] exactly: the same face shape, eyes, eyebrows, nose, lips, jawline, facial hair and skin tone. His hairstyle and clothing match [Image1]. He is clearly recognizable as the same person in every frame, including when he smiles, laughs or turns his head.
Lesson 6: resolution and length
At 480p the face in this clip is only a few dozen pixels wide, so fine details are limited. HD output keeps more facial detail.
Longer clips drift more. A 15-second version tends to stay closer to the photo than a 30-second one.
Generation takes time: a 15-second clip usually takes around 10 minutes, a 30-second clip longer. Tell users they can close the page and get an email when it is ready, otherwise many will think it is broken.
A note on responsible use
Only use photos of yourself or of adults who agreed to it, and do not use photos of celebrities or public figures. Label AI videos as AI when you post them. The original freestyle and the song belong to Bulin 47 and Kevin Wai: if you like the track, stream "Dame Hombro" on the official platforms.
If you want to see these ideas in practice, our [Bulin 47 AI video generator](https://videobulin47.com/) takes one photo and returns a vertical 15- or 30-second MP4 with the original sound plus a silent copy. It is an independent fan tool, not affiliated with Bulin 47 or Kevin Wai. Credit packs are one-time payments (new users can make their first video for $4.99), and failed generations are refunded automatically. There is also a short [step-by-step guide](https://videobulin47.com/how-to-make-bulin-47-ai-video) with photo tips.
If you are building something similar, I hope these notes save you a few test runs.