From Step Video to a working intro

Theta EdgeCloud has had a text-to-video service called Step Video for a while. I'd never run it. This week I did — twice — and then turned the output into an intro I can use in my weekly videos.

Step Video is straightforward: write a text prompt, get a short clip back. The model runs on EdgeCloud, same as the other inference services. Remotion is what I've been using to build the weekly videos themselves — programmatic video composition, React components that render to MP4. Step Video gives you raw clips. Remotion frames, times, and composes them.

The first test

Started with a 3-second job — small, cheap, calibration only. The output is stylized abstract motion, which is what the prompt asked for.

• Frames: 72 (3 sec at 24 fps)
• Duration: 15 min 42 sec
• Cost: $0.20
• File size: 72 KB
• TDROP rebate: ~$0.01

Worked. But 3 seconds is too short to function as an intro — there's no time for a logo to land before the clip ends.

Going longer

Same prompt, same seed, doubled the frame count. I wanted the style identical and just more of it.

• Frames: 144 (6 sec at 24 fps)
• Duration: 70 min 4 sec
• Cost: $0.20
• File size: 233 KB
• TDROP rebate: ~$0.01

Three observations I didn't expect

Same cost regardless of length. Both jobs landed at $0.20 flat. Per second of video, the 3-second job cost $0.067; the 6-second job cost $0.033. Longer generations are noticeably better value per second of output. Generating one 12-second clip would be dramatically cheaper than four 3-second clips of the same length.

Time scaled non-linearly. Doubling the frames didn't double the wait — it took roughly 4.5x as long (15:42 → 70:04). I don't know exactly why. Theta's API doesn't differentiate between "queued waiting for a node," "actively rendering," or "post-processing." It just reports "processing" until the job is done. Possible causes: nonlinear compute scaling with frame count, node availability fluctuating between the two runs, queue depth I couldn't observe. Two runs isn't enough data to separate any of those.

File size scaled to 3.2x, not 2x. 233 KB vs 72 KB. Doubling frames produced 3.2x the bytes. Possible causes: higher bitrate on longer generations, more motion complexity in the output, different encoder settings. Same caveat as above — not enough data points to say.

Turning it into an intro

Step Video gives you a clip. To make it a recurring weekly intro I needed three things on top: my logo placed over the clip, a fade-out to black so the cut into the rest of the video is clean, and versions sized for the platforms I publish to.

Remotion handles all of that. One reusable Intro component takes width, height, and fps as props — same source file in, different canvas dimensions out, using object-fit: cover so there are no black bars regardless of aspect-ratio mismatch. I rendered three versions for the formats I publish to (vertical for Reels/TikTok/Shorts, horizontal for YouTube/X, square as a reference). Here's the 16:9 cut:

What I'm taking away

I have an intro I can use in next week's video. Total Step Video cost: $0.40. Remotion was my own time.

For anyone considering Step Video for similar work: the cost model rewards longer generations. Wait times aren't predictable from one run to the next yet, so don't plan anything time-sensitive around Step Video output. The output is good enough for stylized intros, b-roll, and abstract backgrounds. It is not good for anything requiring precise content — the model doesn't render text or logos reliably. Combining it with a post-processing layer like Remotion is how you get production-ready footage.

Two data points isn't a benchmark. The non-linear time and size scaling might smooth out across more runs, or it might get worse. Right now this is what they were.

— Jacob