Building an AI sci-fi serial: a 5-episode workflow with Flux, Kokoro TTS and pitfalls

darkcashxtr · reddit · 2026-10-08

A creator shares the full pipeline behind his AI-assisted sci-fi serial (narrated slideshows following humanity across 1,000 years from 2100), after 5 episodes:

Pipeline: prose co-written with Claude using tags like [NARRATOR] and [PRIME] to map lines to voices; Kokoro TTS with one voice per tag; one Flux prompt per script beat (episode 5 alone has 381 images); editing synced via Python script and CapCut.
What broke: Flux garbles any text inside images (screens, signs, labels), so frames are reviewed and dropped; episodes neared 50 minutes and were split into two 20-minute parts.
TTS findings: Zonos 2 was a disappointment (voice tags made it sound drunk); Qwen3-TTS matched Kokoro at its best but derailed with emotional tags and seems to only work well via their cloud platform; Kokoro remains the only viable local option, albeit flat.
Open problems: adding real emotion to local TTS (shaky voice, nervous laugh, whisper) and keeping characters visually consistent across hundreds of generated images.

Original post →

Building an AI sci-fi serial: a 5-episode workflow with Flux, Kokoro TTS and pitfalls

More from Multimodal