MakeHumanAI started as a set of free document tools — photo resizers, PDF compressors, signature makers. This week it grew a second half: a faceless YouTube video creation service. You send a topic; we return a finished video — script, voice, clips, captions, thumbnail.
Here is how the production line actually works, because the stack is the interesting part.
The pipeline
Script. A large open model writes the script in a documentary structure — hook, one idea per scene, plain sentences a voice can carry. A 10-minute video becomes roughly 70 short scenes, each with its own narration line and a visual brief.
Voice. We run text-to-speech through HeyGen's Voice Turbo API while it is free, with a local Kokoro model as the permanent fallback. Every scene is generated as its own audio file, which makes retries and revisions cheap — fix one scene, not the whole track.
Clips, not slides. This was a hard rule from day one: no photo slideshows. Each scene gets real motion footage, matched to what the narration is saying at that second, pulled from free stock libraries through their APIs, then graded with one consistent cinematic look — same contrast, same warmth, a light vignette — so stock from five different cameras still feels like one film.
Captions and assembly. Captions are chunked to short phrases, timed from the real audio durations, and burned in. Assembly is ffmpeg end to end: per-scene segments, a concat filter join (stream-copy concat drifts timestamps — ask us how we know), a quiet generated music bed under the voice, and title and end cards.
The business model, briefly
The sample on the service page is free and short — about a minute — so anyone can judge quality before paying. Full videos run on subscriptions: $49/month for two videos, $399/year for twenty-four. Payment is crypto-only (USDT or BTC) with a small live calculator on the page that converts the plan price at the current BTC rate, and orders come through an on-site form — no accounts, no checkout platform taking a cut or holding payouts for a week.
What we learned building it
- Verify every stock clip with your eyes. Search relevance is loose; our first sample had a sparkler, a fantasy underwater city and the Great Wall of China in a video about Rome. Tag filters plus a thumbnail review fixed it.
- Generated audio formats vary per scene (sample rate, mono vs stereo). Normalize before you join anything.
- A demo is worth more than any description. The service page converted from "trust us" to "watch this" the moment the sample went in.
The free tools are still the front door. The studio is what is behind it now.
Top comments (0)