Summary
Nick Saraev shows how to build faceless, educational VTuber-style videos using AI, with the video itself narrated by an anime avatar speaking Japanese. The result is a fully AI-generated video that was not produced by a video model but assembled from generated assets and animated with code.
The pipeline starts with character creation using OpenAI’s GPT Image 2.5, producing pixel-art anime portraits with transparent backgrounds, plus variants for mouth shapes, blinks, expressions, and pointing poses. Voices come from Fish Audio’s Dramatic 3 model, which offers a library of over two million voices and cloning from about ten seconds of audio. An AI agent then writes the storyboard as HTML scenes, animates the elements, transcribes the audio with word-level timestamps for subtitles, and renders frames, in this case by offloading the job to an older MacBook.
Costs are itemized: roughly $4.50 for the one-time character set and about 35 cents per finished video. Saraev also explains why stylized characters avoid the uncanny valley, and how accents and a light noise layer help mask robotic voice artifacts. The video is sponsored by Fish Audio.
📺 Source: Nick Saraev · Published October 30, 2026
🏷️ Format: Tutorial Demo
