I have had such a busy week after the 5 day Bootcamp with Fully Focused, that I have decided this week needs to be about decompressing. Apart from attending a few short meetings, my plan is to slow down, make art and reconnect with the parts of creating that I enjoy the most.
Last night, I decided to make a quick music video while testing two new AI video models, MiniMax H3 and ByteDance Seedance 2.5. What began as a small experiment quickly reminded me why I love creating music videos so much.
Both models genuinely impressed me, but for different reasons. Rather than forcing myself to create the entire video inside one platform, I used each tool for the particular thing I felt it did best.
That is something I have learned through years of making AI films. You do not have to be loyal to one platform. Sometimes the strongest workflow comes from combining several tools and allowing each one to handle a different part of the production.
As usual, I started with the story.
Before I generated any movement, planned any camera angles or thought about lip sync, I needed to understand what the music video was actually about.
The song explores the idea of a woman who loves too much. She gives so much of herself that her love becomes overwhelming, both for her and for the person receiving it. She eventually needs to learn about boundaries, control and how to love someone without losing herself.
I represented this idea through a robot woman.
In the video, she is learning how to hold objects without crushing or breaking them. This becomes a metaphor for how she approaches love. She wants to touch and care for her lover, but she has not yet learned how to control her own strength.
I did not want the robot element to feel like a random science fiction concept. Her physical struggle needed to reflect her emotional struggle. She is not trying to hurt anyone. She simply has not learned how to handle something fragile.
That became the emotional foundation for the entire video.

I created the original image of the robot woman using Midjourney.
This image became the visual starting point for the music video and is also the first shot you see when the video begins. Establishing a strong opening image helped me decide what the lighting, character design and general atmosphere of the video should feel like.
When I create a music video, I normally try to make the first image strong enough to communicate the world before the character has even moved. The audience should immediately understand something about the mood.
Once I had that image, I could use it as a reference across the other video tools. This helped me maintain some level of consistency while still experimenting with different models.


The song was created in Suno. I used my own voice as a Persona, although I can barely hear myself in the finished song.
Before generating the lip sync scenes, I divided the track into sections of approximately 15 seconds. This made it easier to work with the generation limits of the different video models. I vibe coded a simple app using Claude code as a HTML document which I have stored on my computer to help me cut the tracks up without me having to do it myself

Breaking the song into smaller pieces also gave me more control. Instead of asking the AI to understand an entire song, I could focus on one performance, lyric or emotional moment at a time.
I uploaded each section of the music into Hailuo H3 and used the audio as a reference. This process can work with a complete section of music or with an isolated vocal track. Using an a cappella can sometimes make the vocal performance easier for the model to understand, but it is not always necessary.
My prompt for the opening performance was extremely simple:
“Camera very slowly pans out as she sings with no emotion.”
I did not overload the prompt with unnecessary instructions. I knew I wanted a controlled performance, a slow camera movement and very little emotion in her face. That restraint suited the robot character and gave the scene a slightly unsettling quality.

The new MiniMax H3 model was the tool that surprised me the most when it came to lip sync.
H3 can work with text, images, video and audio as part of the same generation process. It can also produce clips of up to 15 seconds with native audio support.
For this video, I mainly used it when the robot woman needed to sing. The mouth movement connected naturally with the music, and the performance felt more believable than many of the lip sync tools I have previously tested.
Technically, I could have attempted to make the entire video in one program. However, I do not believe in limiting a project just to prove that one tool can do everything.
At the moment, the lip sync I achieved with Hailuo H3 was unmatched within this particular workflow. That made it the right tool for the close ups and performance shots.

I also used Hailuo’s Storyboard feature to help develop the surrounding scenes.
I uploaded my original character image and used the tool to explore how the story could move beyond the opening performance. It helped me turn the central idea into a series of visual moments.
The storyboard stage gave me something more practical than a collection of random beautiful images. It allowed me to think about the order of the scenes, what the robot was learning and how her relationship should progress visually.
AI can generate an impressive shot very quickly, but an impressive shot is not automatically a story. The storyboard gave the images a reason to exist.
Once I was happy with the scenes, I took several of the storyboard images into Seedance and animated them there.

I used ByteDance’s Seedance 2.5 for many of the cinematic story scenes.
Seedance 2.5 is designed around longer visual storytelling, reference control and editing, with support for sequences of up to 30 seconds.
What impressed me was its ability to understand what was happening inside the frame. It gave me useful movement without making every character or object move for the sake of it.
For the scenes where the robot interacts with objects, I needed the action to be clear. The audience has to see that she is trying to be gentle, failing and slowly learning to control herself.
Seedance allowed me to focus on those small physical actions. These scenes were less about singing and more about visual storytelling.
I also liked being able to take an image created during the storyboard process and approach it as a proper shot rather than simply turning it into a moving picture.

I used Kling AI for additional clips where I wanted a different interpretation of the movement.
Sometimes I will run a similar scene through more than one model because each model notices something different. One might produce stronger body movement, while another gives me better camera motion or a more convincing facial expression.
Kling has been part of my workflow for a while, so I already understand how to prompt it and what types of movement I am likely to get from it.
I selected the Kling generations that supported the existing story rather than including them simply because they looked impressive. Every shot still needed to belong in the same emotional world.
Another important part of this workflow was creating my own small production tools with Claude.
Using Claude, I vibe coded an FPS fixer and an audio cutter.
The audio cutter helped me divide the song into the smaller sections I needed for the video generations. Instead of repeatedly cutting every piece manually, I could use a simple tool created specifically around my own process.
I also created an FPS fixer to help deal with clips that were generated or exported at different frame rates. When you combine videos from several AI platforms, small technical differences can create problems inside the final edit.
A clip might play strangely, fall out of sync or feel less smooth when it is placed next to footage created somewhere else. Having my own FPS tool gives me a faster way to prepare the footage before editing.
I am not a traditional software developer. However, vibe coding has allowed me to create small solutions for problems I regularly experience as a filmmaker.
The tools do not need to become huge commercial applications. Sometimes the most useful piece of software is something simple that removes one annoying task from your personal workflow.


I brought all of the generated footage into Filmora for the final edit.
This is where the different models stopped feeling like separate experiments and began to feel like one complete music video.
I matched the shots to the music, adjusted the pacing and selected the clips that communicated the story most clearly. I was not trying to use every generation. I only included the footage that strengthened the final piece.
The edit is also where I could control when the audience saw the robot singing and when they saw her trying to interact with the physical world.
The performance scenes carried the emotion of the song, while the story scenes gave that emotion a visual meaning.
The biggest lesson from this AI music video creation process is that there is no need to search for one perfect platform.
Hailuo H3 gave me the strongest lip sync for the performance. Hailuo’s Storyboard tool helped me develop the scenes. Seedance 2.5 handled much of the cinematic storytelling. Kling gave me extra movement options. Midjourney created the original visual identity. Suno produced the song, and Filmora brought everything together.
Claude also allowed me to build practical tools around the workflow instead of waiting for someone else to solve those technical problems for me.
Each platform had a specific job.
Overall, I am really impressed with how the video came out. More importantly, making it has started to relight my passion for music videos.
After such a busy week, I needed to create without pressure. This small test became a reminder that experimentation does not have to be meaningless. You can test new technology while still beginning with character, metaphor and story.
The technology continues to change, but the story is still the part that gives everything else a purpose.