How I Made a Blockbuster AI Short Film on a $125/Month Subscription

No crew. No studio. No film school. No permission.

I'm DAJAI.IO — an independent hip-hop artist from Las Vegas. On April 13, 2026, I released SIMULATION: A DARK Library Film. It's a 7-minute narrative short film about a man who discovers he's the only conscious being inside a simulation, and that the music he created is the only thing the system cannot replicate.

The entire film was made with AI video generation tools on a $125/month Higgsfield Creator subscription. Total spend across two months of production: roughly $200.

This is not a tech demo. This is a story. And this is exactly how I made it.

Why I Made This Film

SIMULATION is part of the DARK Library — an audiobook-album hybrid series I created that transforms classic texts into sonic experiences. Three albums dropped in ten days:

• TOO DARK: The Point of No Return — March 29, 2026
• DARK I: Outwitting the Devil — April 7, 2026
• Simulation — April 8, 2026

The Simulation album explores the idea that reality is a constructed layer and music is one of the few things that breaks through it. The album came first. Then I realized the concept was too visual to stay audio-only. The album became a film.

I didn't plan to make a movie. The music demanded it.

The Tools

Everything was done through Higgsfield's platform on their Creator plan at $125/month. Here's the full stack:

Video Generation Models

Seedance 2.0 (ByteDance) — The primary engine. Seedance consistently produced the most cinematic output. The motion quality, lighting comprehension, and temporal coherence were leagues ahead of what I'd seen six months earlier. About 70% of the final film is Seedance shots.

Kling 3.0 (Kuaishou) — Used for secondary shots and when I needed a different motion style. Kling excels at slower, more deliberate camera movements. I used it for the ambient sequences and some of the close-ups.

Veo 3.1 (Google) — Environmental and atmospheric sequences. Veo handles wide shots and environmental lighting better than anything else I tested. The opening frequency test sequence is Veo.

Cinema Studio 3.0 — Higgsfield's own scene composition tool. Used for assembling multi-element shots and adjusting composition after generation.

Character Consistency

Soul ID — This is the critical piece. Without consistent character faces across shots, you don't have a film — you have a slideshow. Soul ID let me lock character references so that DAJAI.IO, Miko Melts, BB Monroe, and Solana Conejo look like themselves in every single frame.

The workflow: generate a hero image of each character using Soul model with detailed prompts, lock it as a Soul ID reference, then use that reference for every subsequent generation featuring that character.

Post-Production

DaVinci Resolve — All assembly, color grading, sound design, and export happened in Resolve. Free version. The color grade was critical — I needed visual cohesion across shots generated by three different AI models with different color science.

Pre-Production: Character Lock

Before generating a single video frame, I spent two days on character design.

The Cast

• DAJAI.IO — The Architect: The only conscious being in the simulation. Knows something is wrong.
• Miko Melts — The Loop: The simulation's first attempt at connection. Beautiful, present — but she repeats.
• BB Monroe — The Inversion: Everything about her is slightly wrong — the geometry is off, the timing is uncanny.
• Solana Conejo — The Signal: She never speaks. She slides headphones across a desk. She is the only one who is real.

Each character got a Soul ID hero image. I generated 15-20 variations of each and selected the one that felt most like who the character needed to be. Then that image became the immutable reference for every shot.

Production: Shot by Shot

The Multi-Model Strategy

Here's what nobody tells you about AI filmmaking: no single model is best at everything. Seedance does cinematic motion beautifully but sometimes struggles with hands. Kling handles slow camera movements better. Veo does environments that the others can't match.

I treated each model like a different lens in my camera bag. The right tool for the right shot.

Credit Management

• Average cost per usable shot: 3-5 generation attempts
• Total generations attempted: ~150
• Shots in final film: 40+
• Success rate: About 30% of generations were usable

Prompt Engineering for Narrative Film

AI video prompts for films are fundamentally different from prompts for standalone clips. You need:

1. Consistent lighting language — I used "warm tungsten interior light, soft shadows, 35mm film grain" as a base for every indoor shot
2. Camera language — "slow push-in," "static medium shot," "handheld close-up" — be specific about the camera behavior
3. Emotional direction — "contemplative," "uneasy," "intimate but distant" — the model responds to emotional cues
4. Negative prompts — "no text overlays, no watermarks, no sudden camera movement" — tell it what NOT to do

The Hardest Shots

Shot 13 — Solana's hands sliding headphones across the desk — took five attempts. Hands are still the hardest thing for AI video. The solution: generate the hand motion separately from the face, then composite the best of each in Resolve.

Post-Production

Color Grade

Three different AI models means three different color palettes. Without grading, the film looks like a compilation, not a story. My approach:

1. Base LUT: Custom LUT with crushed blacks, warm midtones, slightly desaturated highlights
2. Per-shot corrections: Matched skin tones across models, adjusted exposure for consistency
3. Scene-specific looks: Miko got cooler clinical grade. Solana got warmer. The dissolution went hyper-saturated before cutting to black.

Sound Design

Beyond the album tracks, I added room tone for spatial depth, transition effects between scenes, and a low-frequency hum that builds throughout — the simulation's heartbeat.

Total Cost Breakdown

No, I didn't use a $4,000/month enterprise plan. No, I didn't have a team. No, I didn't go to film school. I had a MacBook, a subscription, and a story that wouldn't let me sleep until I told it.

What I'd Do Differently

• Start with the shot list. I went in with a loose story and figured out shots as I went. Next time, every shot planned before generating.
• Dedicated character wardrobe. Soul ID locks the face, but clothing varies. Be more specific about wardrobe in every prompt.
• Generate at higher resolution from the start. Upscaling early low-res shots shows.
• Budget more credits for hand shots. Hands are genuinely difficult. Allocate 3x credits for any shot involving hand interaction.

What This Means

A year ago, making a narrative short film required tens of thousands of dollars, a crew, locations, permits, and months of post-production. Today, one person with a laptop and a story can produce something that stands next to traditionally-produced content.

I'm not saying AI replaces filmmakers. I'm saying it removes the barriers that kept most people from ever becoming filmmakers in the first place.

I'm a rapper from Las Vegas. I make music on sovereign AI infrastructure I built in my apartment. And now I make films too.

The simulation is cracking.