Directing a Full AI Film Crew Solo: Step-by-Step Short Film Production with Gemini 4 + Omni 1.2
Directing an Entire AI Film Crew Solo: End-to-End Short Film Pipeline with Gemini 4 + Omni 1.2
Complete 7-step production playbook and 10 ready-to-run prompts by @leo_xiaolei: talk only to Gemini 4 in Antigravity while it writes comedic multi-camera scripts, locks cross-shot character fidelity, automates Google Flow (Omni 1.2) over Chrome CDP port 9222, redubs multi-character lines with local 48kHz VoxCPM2 + Whisper millisecond lip-sync, and renders pixel-sharp CCTV HUD overlays and early-reflection masters via Python (PIL) + ffmpeg.
1. Four-Tool AI Crew Architecture & Multi-Camera Script Engineering
Generating isolated 10-second clips in Omni 1.2 is the easiest part of AI filmmaking. The real production bottlenecks are structural comedy writing, cross-shot character/wardrobe drift, native video TTS accent/polyphone errors and clipped final syllables, and garbled in-model HUD text.
Instead of juggling a dozen web tabs and timeline editors, you act purely as Executive Producer talking to Gemini 4 inside Antigravity. Gemini 4 orchestrates Omni 1.2 (cinematographer & actors in Google Flow), local VoxCPM2 48kHz MLX + DeepFilterNet3 + mlx-whisper (multi-role voiceover & millisecond lip-sync QA), and Python PIL + ffmpeg (HUD rendering, room reverb, and lossless mastering).
- Multi-Camera Episodic Structure: Use 5 fixed high-angle (35-degree corner wide-angle) CCTV shots—including 1 monochrome IR night-vision shot for pacing contrast—joined by 0.12s-0.15s static click transitions to eliminate continuous motion-matching failures.
- Physical Comedy Contrast & Model Guardrails: Design every 12s shot around "solemn verbal claim -> betraying hand gesture -> bystander callout," cap each shot at <=3 characters with strictly 1 speaker at a time, finish dialogue before 10.5s (+1.5s reaction hold), and paste an identical ~30-word character appearance string across every prompt.
2. Chrome CDP Automation, Local VoxCPM2 + Whisper Lip-Sync & Python HUD Mastering
Launch Chrome with remote debugging enabled (port 9222) logged into Google Flow so Gemini 4 submits all 5 prompts sequentially, archives versioned takes (raw_scene_01 through _v2/_v3 without overwriting), extracts 1 fps audit frames, and pads rushed openings/endings via ffmpeg tpad (+0.35s head clone, +0.45s tail clone).
To fix native video TTS slurring and polyphone errors, strip vocals while retaining lowered ambient room noise, synthesize 48kHz character voices in VoxCPM2 with trailing ellipses ("...") for a natural +0.25s breath decay, run two-pass mlx-whisper verification protecting the 50ms silent occlusion before affricates/fricatives (plus +0.26s tail fade), compress inter-clause silences (>0.25s -> 0.14s) with 1.05x-1.15x atempo lip-sync alignment, blend early-reflection reverb (aecho=0.85:0.82:22|48:0.14|0.07), and draw 100% crisp 1920x1080 CCTV brackets, blinking REC timecodes, comedic AI tracking boxes, and Whisper-timed subtitles in <200 lines of Python PIL.
Run this short-film project using our "Lead Director + Independent QA" workflow until final master delivery:
1. Persistent Project Rules Checklist: Compile all creative and technical rules into a project checklist file and verify against it at every stage:
- 5 independent fixed high-angle CCTV camera shots (12s each, <=3 characters per shot, 1 monochrome IR night-vision shot in Scene 4).
- Verbatim 30-word character wardrobe/appearance lock pasted identically across every Omni 1.2 prompt.
- Strictly 1 speaker per timestamp window ("Only Character A speaks clearly; B and C keep lips closed"); all dialogue ends before 10.5s with a 1.5s silent reaction hold; pure clean video output with zero in-model text, subtitles, or UI borders.
2. Automated Google Flow Execution (Chrome CDP): Connect to my Chrome remote debugging session, set Video mode (x1, Omni 1.2, 16:9), generate and archive versioned clips (_v2, _v3 without overwriting), and extract 1 fps frames via ffmpeg to inspect head/tail stability.
3. Local Multi-Role Redubbing & Millisecond Lip-Sync: Synthesize 48kHz role-specific vocals via VoxCPM2 (append trailing ellipses to prevent tail clipping), run two-pass mlx-whisper verification (preserve >=50ms pre-affricate occlusion and +0.26s tail decay), tighten >0.25s pauses to 0.14s with 1.05x-1.15x atempo alignment, and mix early-reflection palace reverb (aecho=0.85:0.82:22|48:0.14|0.07) over ducked ambient room tone.
4. Code-Driven Post-Production & Fresh-Eyes QA: Clone 0.35s head / 0.45s tail frames via ffmpeg tpad, draw crisp 1920x1080 CCTV corner brackets, blinking REC camera labels, frame-accurate timecodes, comedic tracking boxes, and Whisper-synced bottom subtitles in Python (PIL/numpy), join cuts with 0.12s static clicks, and audit the final master at 2 fps + 3 frames around every cut.Engineering Comparison Matrix: Anti-Patterns vs. Production Best Practices
| Dimension | Common Anti-Pattern | 2026 Production Standard |
|---|---|---|
| Camera & Character Consistency | Complex camera pans/zooms with paraphrased outfit descriptions causing edge warping and character drift | Fixed 35-degree corner wide-angle shots + verbatim ~30-word character appearance string pasted across all prompts |
| Dialogue Timing & In-Model UI | Packing lines up to 11.5s and asking the video model to draw CCTV timestamps/subtitles (clipped endings & garbled text) | End lines before 10.5s (+1.5s silent reaction hold), enforce pure clean frames, and render 1920x1080 HUD/subtitles via Python PIL |
| Voiceover & Lip-Sync Fidelity | Relying on native video TTS or naive silence trimming that chops 50ms pre-affricate occlusions and drifts off-lip | 48kHz VoxCPM2 with trailing ellipses + two-pass mlx-whisper boundary guard (+0.26s tail) + 0.14s pause tightening & aecho reverb |
Pre-Flight Production Verification Checklist
- 01Split script writing and script review into two distinct Gemini 4 personas ("Creator" vs. "Harsh Comedy & Technical Director") before storyboarding.
- 02Never overwrite raw video takes—archive _v2 and _v3 files and use ffmpeg tpad (+0.35s head / +0.45s tail) instead of rerolling minor edge-frame flaws.
- 03Click the toggle button above to inspect all 10 original bilingual prompt templates, 4 production tables, 5 original CCTV/pipeline exhibits, and the interactive AI Film Crew Parameter Studio.
Found this actionable? Copy standard citation & summary
Formatted with canonical attribution link and executive summary—ready for Slack, Notion, or LLM context.
Put Into Practice: Companion Browser Workbenches
Side-by-Side Document & Prompt Diff Tool
Compare PDF reports, system prompts, and localized copy side by side with zero uploads.
One IP · Purity & AI Node Telemetry
Inspect residential vs. datacenter ASN, fraud risk score, and AI service routing reachability.