I built a tool that turns scripts into 30s docs

I’ve been messing around with Claude to automate video creation for a while now—I even set up a pipeline to auto-post to TikTok—but I wanted to see just how fast I could push the generation speed. After testing out Nano Banana 2 Lite last night, I actually managed to get the turnaround time down to about 30 seconds for a short-form video.

The technical stack is a bit of a Frankenstein’s monster, but it works. I’m using GLM-5.2 via Fireworks to handle the scriptwriting and image prompting. For the visuals, it’s Nano Banana 2 Lite, and I’m using gpt-4o-mini-tts for the voiceover. I use ffmpeg to stitch everything together and apply a Ken Burns zoom effect. Honestly, I’m still struggling to get rid of a slight shake in the zoom effect, but it's getting there.

The biggest bottleneck wasn't actually the AI models, but the video compilation process itself. To stop the server from choking, I had to throw a massive 64 vCPU EC2 instance at it just to speed up the rendering.

The most fascinating part for me is the unit economics. Each short video costs roughly 25 cents to produce. About 90% of that cost is purely the image generation (around 3.3 cents per image). Even with the high cost of running a beefy EC2 instance, the margins are interesting.

It feels like we are rapidly approaching a world where on-demand AI video is a reality. This is just a small demo in a very specific niche, but it shows how close we are to having automated content pipelines that feel almost instant.

Has anyone else experimented with this kind of automated video workflow? I'm curious if anyone has found a better way to handle the video stitching without needing such a massive instance.