GTA-Style Diwali AI Film: 12x10s Scenes, Physics Prompts Fix Video Consistency
AcceptableContest168 · reddit · 2026-11-08
A creator shared a full prompt-engineering breakdown of a GTA6-style Diwali short film made with Seedance 2.5 on Higgsfield.
What broke in single 30s generations: messy zipper close-up, hero standing on a car, final 10 seconds falling apart.
What fixed it:
- Split into 12 ten-second scenes instead of one long clip
- Fixed prompt structure per scene: context, characters, location, format, camera, action, lighting, texture, audio, positive locks
- A physics section in every prompt: weight, braking, door hinges, recoil
- Positive locks ("he stands on the roof floor" instead of "don't stand on the car") plus a flyer lock to keep props in place
- Reference sheets first, then tagged into every scene
The full prompt document is publicly shared.
More from Multimodal
- AI-generated Sunday morning cartoon clip shows off retro cel-animation style — MickeyMau5 · 2026-11-08
- Creating new Midjourney styles with striking retro-illustration results — hewarsaber · 2026-11-08
- AI video reimagines the viral Miami mall alien incident as an action scene — kenL3000 · 2026-11-08
- Recreating GTA III's Liberty City atmosphere entirely with MiniMax video generation — Sweaty-Ad4488 · 2026-11-08
- Open-Source Image-to-3D Still Lags Closed Models, Trellis and Pixal3D Tests Fall Short — Realistic_Rabbit5429 · 2026-11-08
- Underdog Demos Woof, a 32MB Speech Model With 31.7M Params Running Offline in a Browser — emmanuelvivier · 2026-11-08