When your Ai model fucks up. But you set the temp to 2 and the top P to 1.0, this is what happens. It was stuck in a hilarious loop in Ai Studio. I literally just fuck around with ai models and discover weird shit like this. The funny parts are when it's stuck and it knows it. 😂 I hope Gemini forgives me for this experiment.
I wanted to share my latest music video and break down exactly how much work actually goes into making one of these. People often assume you just type one magic prompt and the AI spits out an entire finished video, which, sadly, isn't how it works at all! 😂
I’m a trauma-informed mental health nurse, and on my time off, I make AI art to decompress. I constantly challenge myself to make something better than the last video I did—just like I used to do with my photography. If it's not better than your last art piece, why even try to make art?
Here is a window into the actual work that went into this project:
### The Concept & The Character
The original anime character in the video is a blend of a person I know online (and her avatar) and a person I know in real life. I used her trademark beanie and long blonde hair. Honestly, this whole video is a tribute to the people around me, especially the women I know.
The song is written in Spanish first (not translated from English—I'll post the lyrics in the comments!). The lyrics are about being constantly misunderstood to the point where you are so angry you could smash a TV. It's about trauma and holding everything inside, but the song is intentionally upbeat. It captures that feeling of dancing the line between masking and your trauma.
🤍 Phase 1: The Images (The Base)
*(R.I.P. Whisk. Google just killed off one of my favorite image generators. Well, there goes another Google experiment... poof, gone).* To get started, I created the base character and the original 6 scenes in Gemini using Nano Banana. I asked it to put the character in these vaporwave landscapes, collected a bunch of different scenes, and Nano Banana did a pretty seamless job creating my source images.
* **Formatting:** Sized to 9:16 for mobile format.
* **Post-Processing:** Did some color and aesthetic editing in Prequel before moving them to video.
🤍 Phase 2: The Video (Veo)
I almost always use source images in Veo (either Midjourney or Imagen generations). I usually write short prompts myself, but this time I decided to let Gemini write some of my prompts as an experiment.
You can literally give Gemini a short idea and have it write the highly detailed prompt for you. Always let the AI cook. The videos in Veo turned out 🔥.
**Here is an example of a prompt Gemini wrote that I fed into Veo:**
> *High energy anime style. A dynamic, rhythmic tracking camera captures the blonde female character in a green jumpsuit, gold chains, and beanie performing a flawless, upbeat, and cheerful pop-dance routine. She is standing in a surreal, flooded neon room. In stark, darkly comedic contrast to her perfect, happy dancing, the massive stacks of retro CRT TVs behind her are violently short-circuiting, violently sparking, catching fire, and displaying chaotic, glitching error messages. She remains completely oblivious to the fiery destruction escalating right behind her, smiling brightly and maintaining direct eye contact with the camera as she hits every dance move perfectly. Cinematic vaporwave lighting, vibrant neon reflections on the rippling water, fluid animation, 4k resolution.*
>
I used prompts like this to burn through about 800 credits in Veo, using Veo Lite and generating 3 videos per prompt for variety and consistency. By the end, I had about 12 minutes of raw video footage.
🤍 Phase 3: The Music (Suno)
Fun fact: I actually generated the video *before* the song. Veo just made up random English lyrics and had the character sing them, which obviously didn't match my final track.
I took my Spanish lyrics and went to Suno to make a Reggaeton / Hard Style EDM track. I liked the very first song it generated enough to keep it and not roll for more.
**My Suno Style Prompt:**
*(Pro-tip: I repeat words in my prompts to add emphasis. It works.)*
> *female vocals, reggaeton, HARD STYLE, TECHNO, tropical house, electronic, edm, reggae, j-pop, electro latin, emo, romantic, female falsetto and Soprano vocals, falsetto, falsetto, breathy voice, Jamaican dub, witch house, reggaeton, HARD STYLE, TECHNO, 1990's underground rave, progressive, epic, experimental, trip hop, dreamy atmosphere, emo core, r and b, female vocals, reggaeton, HARD STYLE, TECHNO, tropical house, electronic, edm, reggae, j-pop, electro latin, emo, romantic, female falsetto and Soprano vocals, falsetto, falsetto, breathy voice, Jamaican dub, witch house, reggaeton, HARD STYLE, TECHNO, 1990's underground rave, progressive, epic, experimental, trip hop, dreamy atmosphere, emo core, r and b, female vocals, melodic* >
🤍 Phase 4: The Edit (Filmora)
Now it was time to put it all together using Filmora (a $40/year video editing app that is really great once you get the hang of it).
Somehow, I had to edit 12 minutes of 8-second clips to match a Spanish song to make a cohesive music video. **That part, I can't teach anyone.** You just have to listen to the song and manually adjust the video clips to match the beat and the vibe. That is the most time-consuming part of the whole process.
I was genuinely impressed with Veo's dance moves being right on beat! 😂 You really have to learn the tools and piece them together yourself. I hope this gives people a better idea of how much human effort actually goes into making these.