I Set Out to Test If AI Video Could Sustain a Feature Film and Made Seven

bdylsing · reddit · 2026-09-13

A Redditor set out to test whether generative video could sustain long-form narrative and accidentally made seven one-hour films (C1-C7), starting with Sora 2 a month after launch and evolving from iPhone iMovie to Final Cut Pro plus multiple AI systems.

Key takeaways:

• The hard problem isn't generating pretty clips but production memory: after thousands of generations, someone must track characters, outfits, locations, plot and emotional history.
• The fix: the human became the production memory, feeding the model smaller, controlled subproblems.
• Editing creates complexity; individual generations don't have to.
• The films double as a longitudinal experiment documenting both the story and the author's learning as the tech evolved.

He asks the community: is generative video shifting from a prompting problem to a production problem?

Original post →

More from Multimodal

• Creator Chains Seedream 2.5 with Flux 3 via Luma to Make a 40-Second AI Short — LudovicCreator · 2026-09-14
• A Full Prompt Template for Generating a Photoreal Isometric ARPG Trailer with AI — techhalla · 2026-09-14
• HeyGen open-sources 20+ launch videos as plain HTML that Claude Code can edit — _AustinCalvert_ · 2026-09-14
• Wan3.0 Anime Battle Clip Impresses: Sensible Camera Cuts, Consistent Motion — andrew_n_carr · 2026-09-14
• Greek TTS from 3.5h Data: Deterministic Prompts Fix Speaker Drift, Near-Human MOS — kastnerkyle · 2026-09-13
• CTC-TTS paper replaces MFA with CTC alignment for low-latency dual-streaming TTS — kastnerkyle · 2026-09-13

Multimodal channel →