How to Create Cinematic AI Videos (No-BS Guide)
About this video
Jeff Su argues AI video models are powerful for single 8-second clips — his Darth Vader recreation in Google Flow looks and sounds detailed — but they have no memory between scenes, so extending a clip produces a different-looking character, wrong hand/lightsaber, different voice and changed background. That consistency problem, not raw quality, blocks Hollywood-grade work. His workaround, demonstrated with two Gemini mascot skits, is a four-step workflow: 1) Generate the character as a still image, using Google Whisk (free; he normally uses paid Midjourney) — refining with precise reference enabled, which uses Google’s nano banana model, to change only one attribute like fur color. 2) Create the starting frame in Whisk by uploading the character as a subject with precise reference on and prompting the scene; without a reference and precise reference off, even the same batch produces inconsistent characters. 3) Generate video in Google Flow via Frames to Video, uploading the starting frame and prompting dialogue/action — he uses the paid V3 quality model but says V3 Fast for free users works the same, and generates four outputs per prompt to raise the chance of a usable take; he writes prompts with a custom Gemini Gem loaded with best practices. 4) Fix voice consistency in ElevenLabs Voice Changer by uploading each clip and converting to the same voice (Malvorex, the monster voice), then in Final Cut Pro detaching the original audio and replacing only the mascot’s lines so human voices stay original, adding ambient effects. He closes noting Sora 2’s Cameo (real people and pets only) and Recut features help continuity but do not replace this workflow, and third-party all-in-one tools like OpenArt still require manual character and audio work.
Key takeaways
- Consistency across scenes — same appearance and voice — is the central bottleneck in AI video, because video models do not remember prior scenes.
- Start from a still: generate the character in Whisk with precise reference, then build each scene’s starting frame from that same reference image.
- Use Frames to Video in Flow and generate four outputs per prompt to improve odds of a usable clip; workflow matters more than any single tool.
- Fix audio separately: run every clip through ElevenLabs Voice Changer with the identical voice, then splice only the character’s lines in an editor like Final Cut Pro.
- Sora 2 Cameo and Recut are useful features within a broader workflow, not replacements for character generation, prompting, and audio work.
Episode details: published October 7, 2025 · 17:01 long. Summary written by Self Growth Videos from the episode itself. Business results and any income figures mentioned are Jeff Su’s own claims, shared for education — they are not typical or guaranteed.
← Previous: 95% of People STILL Prompt ChatGPT-5 Wrong · Next: The Productivity System I Taught to 6,642 Googlers →
All Jeff Su videos · Jeff Su profile · Jeff Su (official site)
