TESTING GOOGLE FLOW: AI Filmmaking R&D (Omni Uncut)

A completely uncut, behind-the-scenes look at adapting to the massive new Google Flow and Gemini Omni video model updates. In this R&D session, I am testing the new architecture for an upcoming 1970s-style high-budget anime project. The goal was simple: execute a complex, multi-character tracking shot through a crowded disco. The reality was a deep dive into prompt engineering, engine hallucinations, and finding the hard boundaries of the new ecosystem. If you are an AI filmmaker trying to figure out how to lock character geometry, prevent the Omni world model from acting as an autonomous director, or use the Flow Agent for multi-camera coverage, this is the raw trial-and-error process. Technical Note: The desktop audio loopback dropped during this screen recording setup. You will not hear the native text-to-speech engine generation during the playback tests, but the visual prompt architecture, model benchmarking, and UI workflows remain fully intact. Chapter Breakdown: 00:00 - Introduction & The New Google Flow Updates 03:00 - The Characters Tab: Building the "Frank" Persona 07:30 - Creating the Agent Antagonist (Nano Banana 2) 15:30 - Establishing the 1970s Disco Environment 19:30 - The "Lazy Prompt" Pre-Viz Test 23:45 - Architecting the C.S.A.C.S. Prompt Structure 32:30 - Veo 3.1 Lite vs. Fast Model Benchmarking 38:00 - Omni Flash Testing & Style Drift Hallucinations 44:00 - Pushing the Flow Agent for Multi-Angle Coverage 53:00 - Hitting the Veo Fast Model 3-Ingredient Limit 54:00 - First-Principles Omni Prompt Cleanup 01:00:00 - Final Thoughts: The Reality of Multi-Model Workflows #aifilmmaking #googleflow #GeminiOmni #veo3 #OmniFlash #promptengineering #generativeai #aianimation