100 custom brand films. $196.10. One person.
The images already existed, which makes them sound easier than they were. 100 modern brands reimagined as 1970s retail stores, each a Wes Anderson-symmetrical composition on 35mm Kodachrome, each one wrestled into shape over multiple prompt iterations across months of work. The films would more than double it. The gallery was live. But the stores were still.
The concept initially was just image generation so I started there. A film is a much bigger undertaking. That followed on after. After I'd convinced myself I could do it. I wanted to see the stores more alive, which is where Luma AI came in.
I began with the video generation in March. I gave up after finding it too frustrating and time consuming. It took me until May to come back to it fresh and finish it.
I've spent my career in communications and 15+ years writing creative briefs for other people to make the thing with. It's taken me until now, approaching 40, that I can both write the brief, make the thing and do the comms I've always wanted to.
This has been a weird epiphany of sorts in 2026 -- I am no longer labelled by my job title or my past, but they all layer on top of each other as lived experiences. The job, underneath the job title and the weird psychological identity wrapped around it, was always the same: work out what a brand actually is and stands for, find the texture and culture of it, know what makes it feel right or wrong. That comes over time: experience, battle scars, pattern recognition and a sense of taste. None of those things were automated. They were what got fed into the tools.
Modern Retro is what happens when those all finally meet. The brand research and selection. The prompt engineering -- multiple iterations per image to get each one right. The copywriting across 100 store pages. The web development: the gallery, the game, the compare tool, the create-your-own tool. The image generation. The video generation. The press page. The distribution. The tools have caught up. That is the transformation we are all nervous, anxious and scared about. But it is also unlocking what I think could be the most creative generation ever. Because really, I have no right as one person to have researched, selected, written, built, prompted, generated, edited, published and distributed all of this -- all for a fraction of what a creative agency would charge.
On the one hand that is disrupting the very industry I built a career in. On the other it has been the most liberating, freeing, strange, discombobulating and identity-releasing thing I've done. Because if I can do all this, what does that make me now? Who and what am I? Where does that best fit? That all remains unanswered. The honest answer is, I don't know yet.
What the films are
Every brand got a custom cinematic brief. Not a template. Not "generate a video of the storefront." A specific brief, written by hand, for each brand individually: what the camera does, what the light looks like, what is on the shelves, what the atmosphere feels like.
Anthropic got a gentle push past whiteboards covered in equations and stacked research papers. Palace got a dolly into a Soho skate shop where the PALACE neon sign glows red above the counter. Coinbase got a teller window with a brass rates board. Little Moons got a pastel Japanese sweet shop with a glowing pink crescent moon sign above the display case.
The model is Luma AI Ray 2. The format: 16:9, looping, 5-10 seconds. The aesthetic: the same warm 1970s atmosphere as the images. Tungsten light, film grain, shallow depth of field. Each film is a world you want to step into.
The process
I submitted the initial 100 briefs in batches. Within hours I had 97 videos. Three needed a separate run. This is the easy part of the story.
What followed was less clean. 41 of the 100 needed redoing after the first pass. Not because the AI failed, but because I was not satisfied with what it produced. Wrong focus. Watermarks that had crept in. An uncanny human doing something strange. Text that had garbled.
I ran the redo process in five passes, working down from 41 flagged videos to 32 to 18 to 3 to a final handful of special cases. Each pass required reviewing every video, writing new briefs, resubmitting, waiting, reviewing again.
Every version had to be watched in full, critiqued, and analysed against a single question: is this right? Not is it good enough. Is it right. Some took six or seven attempts to get there. By the third or fourth attempt on a single brand I was questioning whether it was worth the time -- I had 100 to do, not one. But something kept pushing me through it.
That last problem -- garbled text -- turned out to be one of the most frustrating in the project.
What AI cannot do
Luma Ray 2 is remarkable at almost everything. Movement, atmosphere, lighting, depth. What it cannot do reliably is spell multi-syllable brand names in neon signs.
Little Moons. I tried six different approaches: different concepts entirely, different camera angles, different ways of describing the sign. Every single generation came back with the same garble: LITLE, LITETE, something that was visibly trying and visibly failing.
This is not a prompt issue. It is a model limit. Short words (REKT, KITH, HUEL) render reliably. Longer names do not. It is frustrating -- but in a strange way, comforting. This still requires human intuition, experience, craft and taste to get right. The tool cannot replace the person who knows what right looks like.
That took seven attempts to get right and it is now one of the best films in the collection. Seven attempts for one film, with ninety-nine more to make, is either taste or madness. They might be the same thing. You feel the wrong version before you can name it, and you do not stop until that feeling is gone.
What you can fix
Some problems didn't need regenerating. I removed watermarks with a single-pass blur filter on the clean CDN source -- a few videos had them creep in and that fixed it.
For Netflix, the issue was different. The original had the right concept: NETFLIX in neon above the entrance, warm light, the right era. But a person turned their head to face the wrong way for their body in a way no filter could fix.
So I extracted the first frame: the person back-to-camera, the sign symmetric and glowing, everything still and correct. Then I applied a Ken Burns slow zoom over five seconds. The camera pushes in. The film grain breathes. What was a still photograph becomes a film. No weird neck and body movement. Just a warm, symmetrical neon-lit entrance to a 1970s video rental store.
That fix probably took thirty minutes all in. The result is better than any regeneration I attempted.
Some fixes weren't technical. They were about getting the concept right. My original Bumble brief had garbled signs on a matchmaking bureau. A new version tried a florist with yellow sunflowers. It looked beautiful. It was also wrong: Bumble is a dating app. I rebuilt it as a 1970s matchmaking agency: a cork noticeboard covered in portrait photographs and index cards, two armchairs facing each other, warm amber light.
Rekt Drinks needed the brand name visible. I lost the tagline entirely and put REKT alone in a bold red neon sign on the back wall of a dark bar. Four letters. Luma handles four letters fine. It didn't like FOR THE FEARLESS though -- it kept garbling that.
Across 199 tracked generation IDs, I learned which AI constraints are worth fighting and which are not. Garbled multi-word text on a moving sign: not worth fighting. Short text, static position, clean colour: almost always works. Map the model's edges and work around them rather than into them. Once you know the edges, the tool stops being something you fight with and instead becomes something you know how to effectively harness.
Fifteen years of writing briefs for others to interpret. Now the distance between my taste and the final frame is close to zero. No middlemen. No admin layer between the thought and the thing.
This is agency in the personal sense, not the advertising one. The capacity to act, and to see your choices on the screen minutes later. It is liberating. It is addictive.
But the part that compels me sits deeper. Refusing to settle for mediocrity. Refusing to run with the obvious mistake. That is what taste is, to me.
The economics
$196.10. That is the total I paid Luma across the entire project: eleven invoices across March and May 2026, covering the initial batch, the four redo passes, and the monthly subscription.
Under $2 per film.
A traditional production company would charge at minimum $5,000 for a single brand film. A slow dolly through a dressed set, a director of photography, a grade, a sound mix. $5,000 for one. The industry rate for a hundred would be somewhere north of $500,000.
I spent $196.
The economics do not just change what you can afford. They change what you can attempt.
The argument
There is a version of this story I could tell as: "AI generated some videos." That version is technically accurate and almost entirely wrong.
Every brief was written by hand. Every brand got a specific concept: not "generate a 1970s shop" but "gentle push past whiteboards covered in equations, research papers stacked, warm academic lighting, thoughtful research institute atmosphere." That brief came from understanding what Anthropic is, the texture of the brand, not just the name.
The AI rendered. The creative intelligence was mine. That is not a defensive claim. It is the point. Modern Retro exists because someone with taste made 100 specific decisions about 100 specific brands and wrote 100 specific briefs. The tool executed. The authorship was human.
That is the model I am interested in. Not AI as replacement, but AI as execution at a scale that makes previously impossible creative work actually possible.
$196. 100 films. One question per brand: what if they opened in 1975?