You can create AI video entirely from 3D scenes now, and it's absolutely mind-blowing.
I'm genuinely excited about this one. Most AI video today means generating short clips of a few seconds each and stitching them together, and hoping things stay consistent. This time I tried something different. I had Opus 5.5 build the actual 3D world: the environments, the characters, the camera moves, the lighting. Then I rendered the film from that world.
Because everything really exists in 3D, the shots can run as long as I want. The camera can fly continuously through a scene, the characters look the same from start to finish, and every moment can lock exactly to the music. None of the usual few-second limit. It feels like a whole new way to make videos.
This is the music video for my track "AGI". There was no video editor and no hand animation. Every shot is a small script, and the computer renders the whole film from those scripts. I directed, and Claude Code (an AI coding assistant) wrote most of the code.
The story: one human and one AI. The AI copies itself, grabs more and more computing power, and ends up as a planet-sized intelligence. It's told in 12 acts that follow the song.
How it was made:
- The music comes first. A script listens to the song and finds every beat, bar and section change. Cuts land on the beat, flashes hit exactly on the kick drum, and the lights pulse with the music.
- Plan and storyboard. I mapped all 12 acts onto the song and approved a storyboard before anything was built.
- 3D models. Some of the models including the human figure were made in Meshy, an AI 3D model generator. The code poses them, so nothing is animated by hand.
- Building the scenes. Several AI agents built different acts at the same time. Quick low-quality previews of the whole film take about 3 minutes, so I could watch full drafts and give notes. I did two rounds.
- The DLSS 5 pass. DLSS 5 is NVIDIA's new AI rendering tech for games. It adds realistic detail and lighting to each frame. Because it's made for games, it also needs to know how everything is moving, so the engine works that out for every frame. I tried four looks and picked "cinematic". It makes solid surfaces like the figure and the machines look much richer, and it tones down the neon glow a little. It adds almost no render time.
- Final render. The full film in 4K took about 80 minutes to render, with the original lossless audio. A final check caught some black squares flickering in the last few seconds. That was a tiny maths glitch, fixed and re-rendered in 4 minutes.
If you're interested in my creative work, you can find me on YouTube, name: Axinovium - Happy to answer questions!