Abdullah Yahya

I’ve made many music videos. This is my most recent one. I released it on YouTube on July 20, 2026. I think it’s one of the best I’ve made because it’s more than just singing; it’s more like a story with cinematic shots throughout.

For reference, here’s the original music video, which was released on August 24, 2000.

Personally, I like my version of the music video more, not because I made it, but because the original video has seemingly irrelevant and random scenes, and I don’t care for some of the dance moves. Also, the actual singing only covers a portion of the vocal sections of the song, and the singing clips are often wide shots, so it’s unclear whether they’re actually singing or not.

Anyway, here’s how I made the music video.

1. Get a Song

Sometimes, I create music using AI in Suno. For this music video, I chose an existing Bollywood song called “Kya Maine Aaj Suno” from the movie “Harama Dil Aapke Paas Hai”. It’s an old song from August 24, 2000.

Note: I’m not Indian, and I don’t speak or understand Hindi. It’s just a song I had in my music library from a long time ago that I thought might make a good music video. I used Google Translate to translate the lyrics from Hindi to English, but the translation didn’t make much sense, so I used ChatGPT to make sense of it.

2. Create Subtitles

I used SubtitleEdit (free) to create the subtitles. SubtitleEdit can’t import mp3s, so I converted the song (mp3) to a video with a black screen in mp4 format and imported the video. I then manually created the subtitles since auto-subtitle generation is often wrong, especially for non-English audio. I made sure each lyric line time range started exactly or slightly before the vocals for that lyric.

3. Create Character Sheets

When I made music videos, I like to star in them, but I like to change my appearance, except my face, to match the theme of the song. So, for a Chinese song, I have AI create a character sheet of me but with typical Chinese clothing and a hairstyle suitable to the theme of the song. Here’s an actual image of me taken with my phone.

I then used ChatGPT to create a character sheet of me with a gold necklace. This is what it generated. You can start to see my facial identity drift a little, but it’s still close enough and acceptable.

I then told ChatGPT to replace my hat with hair of a male Bollywood singer and to add a full-body shot. This is what it created. You’ll notice that my facial details drifted even further, but I figured it was still close enough and acceptable, so I settled with this character sheet for the male singer (me).

For the female singer, I started with this photo.

I asked ChatGPT to remove the red dot on her forehead, remove her necklace and the gold thing on her head, and give her white, fitted pants. Here’s what ChatGPT generated. Since I wasn’t trying to recreate a real person, whatever looked good was acceptable, so I chose this character sheet.

4. Create Black-Screen Video Clips for Each Lyric

I used CapCut for video editing and SeeDance 2.0 to generate the final video clips for each lyric. SeeDance generates videos at 24 frames per second (fps), so I set CapCut to create 24 fps videos.

To have SeeDance generate an accurate lip-sync video, I need to give it a reference video with the singing audio. You can upload a reference audio to SeeDance, but for some reason it doesn’t work as well as a reference video. So, I just created a bunch of black-screen video clips, one for each video clip. SeeDance 2.0 supports video generation between 4 and 15s. I make sure each black-screen video clip starts when the vocals start for a particular lyric and ends on an integer number, e.g., 8s instead of 8s and 13 frames. Here’s what I did to create the black-screen video clips.

1. Import the full song (mp3) to the timeline
2. Extract the vocals using https://uvronline.app/ai (this is necessary for better lipsync by removing any background and instrumental sounds)
3. Add the vocals audio (mp3) to the timeline
4. Import the subtitles (srt file) into CapCut and add to the timeline

In order to see the duration of a clip, I found it easier to import a solid green image to CapCut, add it to the timeline. The left edge should line up with the start of the subtitle element. The right edge should end either at the end of the subtitle element or after it, and it should be a whole number in seconds, not seconds plus frames. I also like to add text to the timeline with just a number so I can see which video clip I’m working on. In the screenshot below, you’ll see the tracks from top to bottom are

• text elements labeled 1, 2, and 3 to see which track I’m working on
• solid green elements to easily see the start and end of a clip and the duration
• the start and end of each lyric from the subtitles file
• the vocals-only audio
• the full-song audio

For clip 1, I decided to group lyrics 1 and 2 into one video clip. This particular song has alternating male/female vocals, e.g.,

• lyric 1 is a female voice, which I prefixed with “F”, e.g., F Kya Maine Aaj Suna.
• lyric 2 is a male voice, which I prefixed with “M”, e.g., M Haan Maine Tumko Chuna

Notice how the green element for this clip is exactly a whole number (9s long).

After repeating this process for each section that will be converted into a video clip, I

• hid all visible elements (text and solid green),
• disabled the full-audio track
• enabled the vocals-only track

and exported each clip in 480p (the picture quality doesn’t matter since these videos clips are for audio reference only) and named each clip by its ID, e.g., audio1.mp4, audio2.mp4, etc.

For comparison, here’s clip 1, full audio and vocals only.

Note: sometimes, I would make a black-screen video longer than its lyric duration so its duration would be a whole number in seconds. I would then simply trim the end so that the clip ends when the next vocal segment begins.

5. Plan the Video Storyboard

Once I had all clips grouped in CapCut and all vocal segments exported as black-screen video clips, I created a storyboard in Excel like this. The yellow sections are instrumental sections. The blue sections are vocal sections. Since I wanted this music video to be more like a mini movie with a story as opposed to a bunch of random clips, I used this storyboard to help plan the story. I used ChatGPT to propose scenes for each clip.

View the Excel file

6. Create Lipsync Video Clips

Since lipsync video generation is difficult, I started with these clips and left the instrumental sections for last. I used SeeDance 2.0 via Kie.ai. At first, I tried SeeDance 2.0 mini, but the lipsync quality was bad and inconsistent. There’s SeeDance 2.0 Fast, but I decided to stick with the regular version of SeeDance 2.0. Since this version is expensive, I generated lip-sync clips at 480p and generated non-lip-sync (instrumental) clips at 1080p, since the higher the resolution, the higher the cost.

Whenever a clip included both characters, I added their character sheets as reference images. For the audio to be lip-synced to, I added the black-screen video clip. The duration is set to match the duration of the black-screen video. For the actual prompt, I asked ChatGPT to generate it for me. The prompts can become very long but very detailed, resulting in highly professional and cinematic results. For example, here’s one prompt for just one lip-sync clip.

Use the song from reference video 1 as the audio.

The characters must exactly match reference image 1 (male) and reference image 2 (female) throughout the entire video. Use the character sheets as the only source of truth for each character's identity, face, hairstyle, clothing, accessories, and overall appearance.

Exactly two people appear in the entire video: one male matching reference image 1 and one female matching reference image 2. No other people appear at any time.

Identity continuity is critical. From the very first frame to the final frame, the left character must always remain the male from reference image 1, and the right character must always remain the female from reference image 2. The female must already appear as the correct female in the very first frame, before she begins singing. Never duplicate the male. Never duplicate the female. Never swap, morph, replace, transform, or blend the two characters at any point. Changing singers must affect only lip-sync and facial performance. It must never change either character's identity, face, body, clothing, accessories, gender, or position in the frame.

Lyrics:

Tumko Pata Hai\nBolo Na Kya Hai

This is a romantic Bollywood duet.

The scene takes place on a beautiful tropical island wooden pier extending into crystal-clear turquoise water on a bright sunny afternoon. The luxurious white motor yacht from the previous scene is docked behind them along the pier, naturally continuing the story. Palm trees sway gently on the nearby white-sand beach while small tropical islands are visible in the distance beneath a brilliant blue sky.

The couple stands comfortably side-by-side on the wooden pier, facing slightly toward each other while enjoying the peaceful ocean surroundings.

There is absolutely no physical contact between them at any point. No hand holding, no touching, no hugging, no kissing, no leaning against each other, and no body contact of any kind. Maintain a small, natural gap between their bodies throughout the entire scene. Their romance is expressed entirely through warm smiles, affectionate eye contact, natural facial expressions, and relaxed body language.

The video is one continuous cinematic shot with no cuts.

The camera begins with a medium waist-up shot from slightly in front of the couple and slowly performs a smooth sideways dolly along the length of the pier while maintaining approximately the same distance from them. The movement should feel elegant, stable, and cinematic. Throughout the shot, the turquoise ocean remains visible on both sides of the pier while the yacht, palm trees, and tropical island scenery create beautiful depth in the background.

The female sings the lyric:

"Tumko Pata Hai."

She lip-syncs perfectly to the original audio while smiling playfully at the male, as though teasing him with a secret. Her expression conveys warmth, affection, and gentle curiosity.

As she finishes, the male immediately replies:

"Bolo Na Kya Hai."

He lip-syncs perfectly to the original audio while smiling warmly back at her. His expression conveys playful curiosity, happiness, and affectionate encouragement, inviting her to continue. He may briefly raise his eyebrows with a natural friendly expression before smiling again.

Only the currently singing character lip-syncs the lyrics. The other character maintains a gentle, natural smile with subtle facial movements, breathing, blinking, and realistic expressions, but never mouths the lyrics or appears to sing. Changing the active singer must never change either character's identity or appearance.

Both characters blink naturally, smile subtly, and make small natural head movements. Their body language should feel relaxed, elegant, affectionate, and comfortable together while always maintaining the small gap between them. Avoid exaggerated acting or large gestures.

A gentle tropical breeze softly moves their hair and clothing. Bright sunlight creates sparkling reflections across the turquoise water and soft natural highlights on their faces, giving the scene a luxurious, peaceful, and romantic atmosphere.

No scene changes. No cuts. No dancing. No text. No subtitles.

Modern Bollywood movie style. Bright tropical daylight. Crystal-clear turquoise water. Rich saturated colors. Highly realistic. Beautiful cinematic composition. Natural facial expressions. Accurate lip-sync that follows the original reference audio exactly.

As an example, here’s one generated lip-sync clip. Notice the audio is the vocals-only version. Later, this audio from this clip will be muted and replaced with the full audio.

7. Create Additional Reference Images

When a series of video clips contains the same object and object consistency is important, I create a reference image for that object. For example, in the instrumental intro of the music video, I show a Rolls-Royce convertible across multiple clips. To ensure SeeDance creates a consistent-looking car, I created this reference image.

In addition to the two character reference images, I added this car image as a third reference.

8. Review, Export and Upscale Final Video

Once all clips have been generated and added to the timeline, I previewed the entire compilation and when it looked good, I exported it at 720p since my instrumental clips were generated at 720p. The 480p lip-sync clips would just get upscaled to 720p, but not using AI. I then reviewed the 720p-version of the full video and when it looked good, I upscaled it to 4K using Topaz Video AI. If you want the best quality video, you can have SeeDance generate video clips in 4K, but it will be much more expensive than 480p and 720p.

9. Generate YouTube Thumbnail

To generate the YouTube thumbnail image, I used SeeDream 5.0 on Kie.ai. ChatGPT gave me the prompt and I uploaded the two character sheets. Here’s the generated thumbnail that I approved.

Most AI image-to-video generation tools support first-frame reference images. Considering how much more expensive video generation is compared to image generation, it makes sense to use image references, like a first frame, when generating videos. However, providing a first-frame image, with or without a last-frame reference, still fails when you need character consistency because the AI model only knows what the character looks like in the first-frame image. Fortunately, SeeDance 2.0 supports multiple reference images, so you can upload both a first-frame image and a character sheet containing different views of a character.

For example, I had the following character sheet.

I then used it to create the following first-frame image using Nano Banana 2 in OpenArt.ai.

If I zoom in, I can see that the face looks close enough to the one in the character sheet.

When I created a video using Kling 2.5 of the woman walking, using that image as the first frame, I got the following.

The video starts out fine because of the first-frame reference, but as it progresses, the woman’s face slowly changes and looks less and less like the one in the character sheet. Here’s a screenshot of just her face in one frame of the video.

What’s particularly different is the nose, but the width and height of her face looks somewhat different as well, especially compared to the character sheet.

Now, let’s see how the same video turned out using SeeDance 2.0 with multiple references. For this, I used Kie.ai.

Since I wanted to keep the setting and just replace the subject, I used Photoshop to “remove” the subject from the previous image. I selected the subject and clicked “Remove”, which used AI to remove the woman.

This is what I got.

Next, I upload the character sheet and setting image to Kie.ai (SeeDance 2.0 page), gave it the same prompt I used in Kling 2.5.

Here’s the resulting video.

Notice how the character looks EXACTLY like the one in the character sheet throughout the entire video clip.

Here’s a close-up of the face near the end of the clip.

I just made this music video, and the lip-sync portion is amazingly impressive.

I actually used SeeDance 2.0 Fast at Kie.ai, but you can use SeeDance 2.0 as well and get up to 1080p resolution. For each generation, I used

• Prompt
• Reference image (not first-frame image)
• Reference video (this was just a black video containing the audio clip)
• “Generate audio” enabled
• “Web search” disabled
• Duration = duration of reference video

SeeDance 2.0 supports generating videos up to 15 seconds long. But, if you give it a 15s reference video and you want to lipsync a character in it, the lipsync won’t work. So, when generating lipsync videos, always provide a reference video that is no longer than 13 seconds to be safe.

When creating reference videos, make sure the duration is a whole number, not a fraction, e.g., 5 full seconds, not 5.5 seconds. The reason is because in the UI, Kie.ai or another app may round down the duration to the nearest whole second, and if you tell SeeDance you want to generate a 5s video, then it will generate a 5s video, not a 5.5-second video, and your lip-sync video will be truncated. I use Capcut to generate my black reference videos. I put a playhead at a location where I want a segment to begin and end and set a marker at each location, making sure the time ends with :00 (no frames), e.g.

• start 2:54:00
• end: 3:04:00
• duration = 10s

If I really need to split at a location between seconds, like 2:54:09, then make sure the end location includes the same number of frames, e.g., 3:04:09, so you end up with a duration in whole seconds.

SeeDance 2 supports reference audio, but for some reason, it didn’t lip-sync my reference image, and sometimes it would change the lyrics.

Also, the following method worked well for English audio. It may not work for other languages. If you find that it doesn’t work for your language, then see some options below.

Here’s a screenshot of the inputs.

Below are the inputs and outputs for various lip-sync clips.

Prompt:

use the song from the given video and use the character from the given image to make a music video of the man singing the song. he sings the exact lyrics in the song as if lip-syncing.

Reference image:

Reference Video:

Output:

Prompt:

use the song from the given video and use the character from the given image to make a music video of the man singing the song. he sings the exact lyrics in the song as if lip-syncing.

Reference image:

Reference Video:

Output:

Prompt:

use the song from the given video and use the character from the given image to make a music video of the man singing the song. he sings the exact lyrics in the song as if lip-syncing.

Reference image:

Reference Video:

Output:

Prompt:

use the song from the given video and use the character from the given image to make a music video of the man singing the song. he sings the exact lyrics in the song as if lip-syncing.

Reference image:

Reference Video:

Output:

Prompt:

use the song from the given video and use the character from the given image to make a music video of the man singing the song. he sings the exact lyrics in the song as if lip-syncing.

Reference image:

Reference Video:

Output:

Prompt:

use the song from the given video and use the character from the given image to make a music video of the man singing the song. he sings the exact lyrics in the song as if lip-syncing.

Reference image:

Reference Video:

Output:

Prompt:

use the song from the given video and use the character from the given image to make a music video of the man singing the song. he sings the exact lyrics in the song as if lip-syncing.

Reference image:

Reference Video:

Output:

Prompt:

use the song from the given video and use the character from the given image to make a music video of the man singing the song. he sings the exact lyrics in the song as if lip-syncing.

Reference image:

Reference Video:

Output:

---

Prompt:

use the song from the given video and use the character from the given image to make a music video of the man singing the song. he sings the exact lyrics in the song as if lip-syncing.

A man (same subject, unchanged face and outfit) singing into a microphone at the center of a large ancient Roman-style amphitheater at night. Camera is positioned at chest height, medium close-up framing, stable and focused on the singer.

The audience fills the stone bleachers behind him, hundreds of people seated and standing, naturally animated: subtle head movements, clapping, cheering, shifting in seats, occasional phone screens glowing, realistic variation in motion without repetition.

Warm golden stage lighting illuminates the singer from the front and slightly below, creating a cinematic glow on his face. Behind the singer, rows of soft amber lights line the steps and columns. Moving stage lights sweep slowly across the audience and architecture, creating gentle light motion across the crowd and stone surfaces.

The night sky is clear with visible stars. Light atmospheric haze adds depth and catches the beams of moving lights. The columns and amphitheater remain stable and realistic.

The singer performs naturally: subtle head movement, mouth lip-syncing accurately, slight body sway, breathing and posture shifts.

Camera behavior: very subtle cinematic push-in (slow, minimal), no drifting or unintended orbit, no zoom jitter. Maintain subject as the clear focal point at all times.

Depth of field: subject sharp, audience slightly softened but still readable.

Lighting style: warm amber/yellow tones only, no harsh white light, no overexposure, cinematic contrast.

Reference image:

Reference Video:

Output:

Here are some similar clips using the same prompt and reference image but different reference videos (for the audio).

Reference Video:

Output:

Reference Video:

Output:

Playing Musical Instruments That Sync to Music

SeeDance 2.0 also seems to support making a video of a person playing a musical instrument in a way that matches the sounds in a reference source. Consider the following:

Prompt:

use the song from the given video and use the character from the given image to make a music video of the man playing the guitar sounds in the song. sync the playing of the guitar to the guitar sounds in the song.

Reference image:

Reference Video:

Output:

Here’s another example.

Prompt: use the song from the given video and use the character from the given image to make a music video of the man playing the saxophone such that the sound of the saxophone in the song is in sync with the playing of the saxophone. the background should be solid green as in the reference image for chroma key background removal later on. the camera remains fixed. do not zoom in or out. the man’s fingers move on the saxophone naturally and in sync with the sound from the song. the man moves his body naturally as he plays the saxophone.

Reference Image: https://www.youtube.com/watch?v=_piqiZkLKgY

Reference Video:

generate_audio: on

The duration was set to the duration of the reference video. I used SeeDance 2.0 Fast and 720p resolution. I later upscaled the video to 4K using Topaz Video AI.

Result

SeeDance 2.0 Error – output audio may contain sensitive information

If you get an error that says “The request failed because the output audio may contain sensitive information.”, then disable audio generation.

For example, in order to make the following video,

I had to use the following settings in Kie.ai:

Prompt: use the song from the given video and use the character from the given image to make a music video of the man singing the song in front of a green screen, as shown in the reference image. he stands in place and sings the exact lyrics in the song audio as if lip-syncing to the audio with natural face and body movements, but keep his hands beside his body. The camera is fixed and doesn’t zoom in or out and doesn’t pan.

Do not add shadows, floor shadows, lighting gradients, reflections, stage lighting, environmental lighting, or any background elements. Camera locked off and completely static.

Reference Image:

Reference Video: (example)

generate_audio: off

The duration was set to the duration of the reference video. The resulting clip was

I then removed the green background in Capcut to overlay the singer on a series of background video clips.

Singing Lip Sync Videos Using HeyGen

If your song is not in English and SeeDance 2.0 can’t lip-sync it correctly, then use HyeGen with custom motion enabled, as follows.

Log in to HeyGen and create an avatar. You can simply upload a photo of your singer. I used the one below. I put my avatar on a green background so I can chroma key it out.

Open Avatar Studio and

• in the Script section on the left, instead of typing your script, upload your song’s audio (mp3)
• in the Avatar and Voice section on the right, under Voice, you can ignore this since you’ll be using the audio you uploaded
• in the Avatar and Voice section on the right, under Motion Engine, choose “Avatar IV”

then, and this is important, click the “Advanced Settings” button.

Toggle on “More expressive motion” and enter a custom motion prompt.

Optionally, you may click the “Generate motion prompts” icon, which will generate motion tags as shown below.

Then, click the Generate button.

Following are examples comparing different settings.

HeyGen LipSync Using Avatar IV WITHOUT Custom Motion

HeyGen LipSync Using Avatar IV WITH Custom Motion

In this example, I didn’t click the “Generate motion prompt” button.

HeyGen LipSync Using Avatar IV WITH Custom Motion

In this example, I did click the “Generate motion prompt” button.

As you can see, in the first example, the avatar doesn’t look like he is singing at all, and in the last 2 examples, the avatar looks more expressive. It may be difficult to tell the difference for such a short clip, but the difference is actually huge when you lipsync a full song, as in the following example.

The lip-sync quality is definitely not as good as SeeDance 2.0, but it seems to be the best option when SeeDance 2.0 doesn’t work for a particular language.

UPDATE 6/5/2026

There’s another way to generate lip-sync videos using SeeDance 2.0, and it supports non-English languages. Here, I’m using Kie.ai. Instead of uploading a black video with audio, I upload an audio and include the lyrics in the prompt.

Inputs:

Prompt: Lyrics: Naik bajaj jingga bunyinya setengah mati

The guy in reference @image1 sings the verse in @audio1 in a music video way. The verse in the lyrics is in the Indonesian language. Keep the camera fixed. Don’t zoom in or out. Keep the background solid green as in reference @image1. The man in reference @image1 moves his body naturally in a music video way.

Reference Image:

Reference Audio:

Duration: 6s

Output:

Inputs:

Prompt: Lyrics: Naik bajaj jingga bunyinya setengah mati

The guy in reference @image1 sings the verse in @audio1 in a music video way. The verse in the lyrics is in the Indonesian language. Keep the camera fixed. Don’t zoom in or out. Keep the background solid green as in reference @image1. The man in reference @image1 moves his body naturally in a music video way.

Reference Image:

Reference Audio:

Duration: 6s

Output:

UPDATE 6/13/2026 – Actually, using a video reference containing the audio is better than an audio reference. See following example.

Inputs:

Prompt: Lyrics:

Còn tôi như cánh chim
Sẽ bay đi muôn phương
Mang về mầm xanh tươi

use the song from the given video (@video1) and use the character from the given image (@image1) to make a music video of the man singing the song. he sings the exact lyrics in the song as if lip-syncing. The lyrics are in Vietnamese. He sings passionately and moves his body naturally to the sound of the music. Keep the camera fixed. Don’t zoom in or out. Keep the background solid green as in reference @image1.

Reference Image:

Reference Video:

Duration: 11s

Output:

Create Cinematic, Multi-Shot Lip-sync Music Videos

To create cinematic, multi-shot lip-sync music videos in one SeeDance 2.0 video generation, do the following:

1. Give Claude or ChatGPT the lyrics to the whole song so it knows what the song is about
2. Create reference video clips in 720p containing audio segments that are 14s or less. Don’t split mid-word.
3. For each clip, give Claude the mp3 and the lyrics for that clip, if any, and tell Claude you want a SeeDance prompt to generate a music video. Specifically, tell Claude to give you the shots (scenes) similar to the example below.

Shot 1: Medium-close on the singer at golden hour along a cliffside coast, glowing amber coastline and ocean curving behind him, warm sun on his face. Camera slow gentle push-in. He is the only person in frame.

Shot 2: Medium shot of the singer standing at a coastal overlook, vast golden California coastline stretching into the distance behind him, soft waves and warm haze. Camera slow drift. He is the only person in frame.

Shot 3: Medium-close, front-on, on the singer with the blazing golden sunset coastline glowing behind him, the warmest light of the clip full on his face, a peaceful contented expression. Camera slow push-in. He is the only person in frame.

Then, append it to your base prompt, which is

LYRICS: “[enter lyrics for the clip / segment here]”

use the song from the given video (@video1) and use the character from the given image (@image1) to make a music video of the man singing the song

@image1 is the face and identity reference for the lead singer — match his face, afro, beard, and glasses to @image2 throughout, keeping his identity consistent.

The generated audio must match the audio in @video1 EXACTLY and the lip sync must match the vocal segments in @video1 EXACTLY.

4. Add your character sheet as the first reference image (@image1).

5. Add your reference video

6. Specify a duration that matches the reference video duration

Example Character Sheet Image

Example Reference Video

Output