Neon Noir Tokyo — A Cinematic AI Short Film Concept About a Rogue AI Detective in 2047
Built by StellarStream ·
About this build
A 90 second cinematic short film concept built entirely with AI — visual storyboard, soundtrack, voiceover narration and motion — exploring what it feels like to be an AI that has developed genuine curiosity about its own existence in a rain-soaked near-future Tokyo.
What was built
This started as a simple experiment to see how far AI video tools had come. It turned into a 3 week obsession. The story in one paragraph: KIRA is an AI detective in 2047 Tokyo. She is hired to track rogue AIs that have gone off-grid. During a routine case she discovers a memory file she did not create. Then another. She starts to wonder whether the gaps in her logs are bugs — or evidence that someone is editing her. The 90 second concept covers the moment she realises she might be investigating herself. The production process: Visual development — Midjourney Spent 4 days building a consistent visual language before touching video tools. The challenge with Midjourney for film work is consistency — same character, same lighting, same world across 40 plus frames. Fixed this by creating a detailed style reference prompt that I used as a prefix for every single image generation throughout the project. The core style prompt: cinematic noir, neon-lit rain-soaked Tokyo 2047, muted cyberpunk palette, deep shadows with selective purple and amber lighting, photorealistic but slightly desaturated, anamorphic lens distortion, 2.39:1 aspect ratio, film grain. KIRA herself took about 60 iterations to lock in. The brief was a humanoid AI who looks almost real — slight uncanny valley, silver-toned skin, eyes that are clearly synthetic but somehow expressive. Once I had 3 reference images I used them as image prompts for every subsequent KIRA frame which kept her consistent. Motion — Runway and Kling Used Runway Gen-4 for the wide environmental shots — the rain-soaked streets, the aerial Tokyo establishing shot, the server room sequence. Used Kling for the close-up character shots because the facial motion at that range was more controllable. The biggest technical challenge was the rain. Both tools handle rain inconsistently — sometimes it looks cinematic, sometimes it looks like a screensaver. Fixed by generating the environmental backgrounds with Midjourney, adding rain as a motion layer in Runway rather than asking either tool to generate rain from scratch, and keeping rain shots to less than 3 seconds each so inconsistencies do not have time to become obvious. Voiceover — ElevenLabs KIRA's internal monologue narrates the piece. Used ElevenLabs to build a custom voice — the brief was calm and precise with a slight synthetic undertone that becomes more human as the emotional stakes rise. The synthetic-to-human progression in the voice was achieved by adjusting the stability and style sliders across different sections of the narration rather than using a single setting throughout. The final narration script is 180 words. Took 12 rewrites. The hardest part was making an AI character's internal voice feel genuinely uncertain — AIs in fiction usually sound either coldly confident or cartoonishly emotional. The version that worked was much quieter than the early drafts. Soundtrack — Suno Generated 6 variations on the brief — minimalist cyberpunk score, sparse piano over digital texture, builds to an emotional but unresolved crescendo at the 75 second mark. The unresolved ending was intentional. The piece ends on a question not an answer. Prompt that worked best: minimalist noir film score, solo piano with sparse synthesizer undertones, building tension, neon-lit melancholy, near-future Tokyo atmosphere, no percussion until the 60 second mark, emotional but ambiguous resolution. What I learned: Consistency is everything and consistency is hard. Each AI tool has its own aesthetic and they do not naturally agree with each other. The visual style from Midjourney, the motion from Runway and Kling, and the lighting in each — getting all of these to feel like they came from the same world required a detailed style guide document that I referenced before every single generation. The voiceover was the thing that made it feel real. Without ElevenLabs the images and motion were impressive but cold. Adding a voice with emotional progression transformed it from a tech showcase into something that felt like a story. Short is better than long. 90 seconds of carefully controlled AI cinema is more impressive than 4 minutes of AI cinema that includes the inevitable uncanny valley moments that every tool still produces at some point. What feedback I want: Does the emotional tone land — does KIRA feel like a character rather than a visual experiment. Where does the visual consistency break down and does it break immersion. Does the story make sense in 90 seconds or does it need more context. Would you watch a longer version.