Adrian Elton Creative

08/05/2026

For the final peak under the hood as to how I made my unbriefed RACV speculative AI TV ad, I thought I'd share the 'look frames' that define what the different 'actors', 'extras', 'props' and 'locations' look like. Once these are locked down, they can be used as recurring references which means all of the keyframes and video sequences that follow are also locked down. That's pretty significant as you'd otherwise have to describe each character and element, time and again, every single prompt. The impact of that would mean that you'd be subject to constant drift; meaning things would look completely incoherent from shot to shot. Not ideal. Unless you wanted that for comedic effect.

Anyway, if you wanna learn how to do this for real - I can be engaged for a modest king's ransom :)

06/05/2026

Drilling down into how I made my unbriefed RACV 'Antique Roadshow'-inspired AI TVC, I think it's really important to underline how far off the process was from being able to input a magic prompt that spat out a finished spot.

To that point, I spent one entire month trying to get the key shot of the host sneezing and accidentally dropping the priceless porcelain figurine.

God knows how many credits and dollars I burnt through, moving from platform to platform, and model to model, trying to get it?

Because it didn't matter how many times I tweaked the prompts, whether through my fair hand, or via Chat GPT, absolutely nothing worked.

Quite simply, it couldn't simultaneously generate the host experiencing a sudden and uncontrollable sneezing fit, which in turn was the cause of him reflexively drawing his arm up towards his nose, which in turn was the cause of him losing control of his grip on the priceless porcelain figurine and dropping it.

Instead it was one absurd sequence after another, most typically featuring his nose spraying explosively, which was then followed by him violently hurling the figurine down.

As this shot was the crux of the whole TV ad there was no 'near enough' that was going to constitute 'good enough'. It had to look like a terrible and shockingly believable accident.

In the end, I came to the conclusion that this level of nuance was beyond what models like Google's Veo 3.1 and Seedance 1.5 were capable of in January 2026. It was only finally nailed once Kling 3.0 was released at the beginning of February. I ran the self-same prompt, once and once only, and 'voila', it was done.

The relief was palpable, as was the sense of vindication that I was completely on the money that the previous generation of video models weren't able to grapple with that level of behavioural/performance based complexity.

That being the case, I thought it'd be pretty funny to cut together a blooper reel of some of the most absurd sequences that the AI generated on route to finally hitting pay dirt with Kling 3.0.

Process-wise, it really isn't that dissimilar from being a director on-set with actors who may, or may not, nail their lines. Sometimes the mistakes will be unexceptional. But sometimes, the mistakes will reduce everyone to tears.

So too here!

Enjoy :)

04/05/2026

In follow-up to the AdNews article about my AI ads, I wanted to throw the spotlight today on the genesis of the second of my two live-action style AI TVCs.

This one is an unbriefed, speculative spot for RACV home insurance.

I was over at my parent's place, mid-January, and they were watching 'Antiques Roadshow'. As I watched the host delicately handle a priceless artefact, I suddenly thought how shocking it would be if he accidentally dropped it. A nano-second later I thought it would make for a pretty funny ad scenario. And a nano-second after that I thought it would work a treat for an insurance group like RACV, spruiking cover for valuables that leave the home.

Once the circuit was complete I knew it would be a perfect subject for exploring AI video for the purpose of creating TV ads.

To move from a neat idea to a produced ad though, I first had to write the idea up as a script. While the blank paper scares the bejesus out of many people, as the idea was so fully formed it was far less intimidating. That said, there's still a moment of truth, when you need to dive in and fine tune from there.

The process forces you to think through how the story is going to unfold as a series of shots. And whereas feature film writers might end up with a wall of post-it notes to externalise the narrative structure - a 30 second TVC rarely requires that level of complexity.

The first step to bringing it to life with AI was creating 'look frames' for all of the key characters - so the host and young woman - as well as the group of onlookers - followed by the mansion itself, the appraisal table, and finally an extra detailed reference image of the fictitious 17th century porcelain figurine.

There were also closeup reference images generated for the host and young woman's faces. These were created using the Nano Banana Pro image model inside of the Leonardo AI platform.

These were drawn upon to create the keyframes for each shot. These became the first frames of the video sequences that were mainly generated with Google's Veo 3.1 video AI model. The key advantage is that Veo 3.1 can generate the dialogue in the same pass.

Across the production of this spot I experimented with ElevenLabs, Freepik, Leonardo AI and Runway ML.

I edited the final sequences in Adobe Premiere, using Topaz Video AI to upscale before grading it with 'Film Nitrate Convert' which added analogue grain.

While I could have generated the music too, I was committed to doing it sans AI. I wanted to amplify the fuddy-duddy world of the ad with music that evoked the string quartet from Fawlty Towers. I composed and recorded the original soundtrack in my home studio. While the violin, viola, cello and French horns were performed with virtual instruments, I played the bowed double-bass.

It ultimately came together in a way that proved how potent this technology is going to become as it continues to improve & cross the final threshold of photorealism.

01/05/2026

In the spirit of getting under the hood of how I made my speculative, unbriefed, Woojer AI TV ad, I wanted to shine a spotlight on the first step, assuming that you've developed your own concept and written the script.

Basically, in the same way that you'd have a casting agent send through headshots of actors for consideration, before you generate your video sequences, you need to first establish what the characters look like. In some ways you get to be more granular and specific than if you were casting for real.

And then you do the same for the locations and any key props.

Fundamentally you're trying to get high quality reference shots (AKA look frames) that can then be used as references for generating your keyframes which bring the various elements together to represent the various scenarios that have been established, shot by shot, in the script.

These keyframes typically become the first frame of the video sequences that are generated using the keyframe as a locked visual reference.

Obviously getting the characters right comes down to the clarity of your vision and what you'll settle for. You also need to define what the characters look like on different angles. The minimum requirement is a full body image showing what they're wearing. And then a facial closeup.

Once you've got the keyframes sorted, it's time to get out the buttered popcorn, and to dust off the director's chair, because you're about to experience your movie-making debut :)

28/04/2026

As promised, just sharing a bit of the background process for the making of my unbriefed spec AI TVC for haptic audio brand, Woojer.

There's a popular assumption in the peanut gallery that you press a single button, and out pops a perfect piece of AI video. If you're producing an advertisement with a specific script however, then it's far more likely that you'll be writing and refining your prompts many times over before you've generated 'the shot' that absolutely nails it.

As such there are endless misfires that end up on the digital cutting room floor. Sometimes they fail to make the grade for very boring, perfunctory reasons. But sometimes the outputs are profoundly wrong and unintentionally hilarious.

That in mind, I assembled the funniest misfires from my Woojer spot and created this little bloopers reel. These sequences were generated care of AI platform Weavy. The weird thing was that these used the self-same prompt that I used on a variety of other platforms, with the self-same video model, Google's Veo 3.1. For some reason though, on the Weavy platform, it treated my prompts as helpful suggestions to which it added its own interpretive dance. Specifically, all of the explosions and crackling lightning were never asked for or even hinted at. When you realise that these are the AI's independent enhancements, it underscores the exasperation of seeing these sequences pop out of the machine. These ones in particular left me clutching my sides such was the magnitude of their ridiculousness.

Hope it duly amuses 🙂

28/04/2026

Following the AdNews article last week I just wanted to throw the spotlight on the genesis of the first of my two recent live-action style AI TVCs.

This one is an unbriefed, speculative spot for haptic audio tech brand, Woojer, whose products use magnetic actuators to translate sound into physical sensation. So you don’t just hear sounds, you feel them.

The idea was simple: An elderly wife complains about her husband’s persistent back and neck pain. We assume it’s age. But he’s been wearing a Woojer Vest while playing video games - and the haptic feedback creates an illusion so powerful he's convinced that he’s been seriously injured.

The process started - while sitting on the Waverly line train - generating the key art for the couple and their apartment using Nano Banana Pro (via Krea), along with reference imagery for the Vest and a classic PlayStation controller.

From there I created keyframes for each of the shots that I'd written up in the script. These were subsequently brought to life as video sequences.

The footage was created across multiple platforms inc. ElevenLabs, Freepik, Krea, Leonardo and Weavy. Google’s Veo 3.1 video model did most of the heavy lifting. The game changer was being able to generate the dialogue and performance in the same pass, rather than having to stitch it together afterwards.
The final assembly was done manually in Adobe Premiere Pro, then graded using FilmConvert Nitrate to push it away from the hyper-sharp AI aesthetic and into something more analogue, filmic and textural.

The soundtrack is an original, Klezmer-inspired pieces that I composed and recorded in Logic Pro. It even features me playing live double bass.

The biggest challenge using AI for this kind of ad was generating performances where the body language, accents and annunciation, and general physicality were convincing and 'human' enough for the comedic beats to land.

And even though it falls some distance from being indistinguishable from real footage, the fact that it fundamentally worked feels like a genuine inflection point in terms of how TVCs are likely to be produced moving forwards.

23/04/2026

Very exciting to have AdNews Australia publish this story this morning about my first foray into live-action style AI TVCs.

Following on from my ‘AI-Ad-A-Day’ series; over the last couple of months I set out to produce two unbriefed speculative TV spots using an AI-led workflow as the production layer. One was for RACV, and the other one was for haptic audio tech brand, Woojer.

I developed both of them end-to-end, including the original concepts, scripts, edits and soundtracks.

Off the back of the work, I’ve been invited to join new AI film production group, ‘Gunshy AyEye’, that’s being set up by Australian film industry legends, Clayton Jacobson and Jason Byrne.

RACV and Woojer.

03/12/2025

Everyone's a critic...

This is my first ever AI generated TVC.

After I wrote my salty review of the Coke Xmas ad for AdNews Australia, I thought to myself, it's all too easy to be the critic, lobbing pot shots from the sidelines. So I asked myself what I might have done had I been given the same brief.

Now I caution this is the first thought that occurred to me which is not how I'd respond to a real brief.

But as compares to the piss-take approach of the spec 'Liquid Death' ad, I wanted to approach this earnestly. So a family friendly spot with cute animals and lots of Coke livery. But unlike the real Coke ad, I wanted to take the use of AI in a new direction. That meant having an ad that actually had a narrative progression more complex than trucks arriving. Trucks arrived.

I did everything I could, including creating the key concept and script, and then used AI to generate the key frames which were brought to life with Sora Pro 2 & Veo 3.1

While I could have used Suno to create the music, I've got a home recording studio and love creating music. So I composed and played everything. The intro intentionally invokes 'Riders On The Storm', but the funk groove is completely of my own invention. That said, I used AI platform, kits.ai, to transform my sung vocals of the iconic 'Always Coca-Cola' line.

But I'm not really a video guy. So when it came to assembling it all I had to literally relearn Adobe Premiere as it's been one hot minute since I last edited any video.

And if you think this was produced with one golden prompt, think again. For this 10 shot ad, I generated 107 video sequences and 71 stills. These were done via Krea, Weavy and Runway.

And if you're wondering what it cost in AI credits, it was $401.43 AUD. So considerably more than the negligible cost associated with generating one-off ad images. But also nothing compared to the production costs for real animation workflows. And while real animation is the gold standard - it requires a certain tier of client to finance that kind of work.

And how long did it take me? Well the idea occurred on 6 Nov. And although I thought I'd have it done in 24 hrs, here we are a month later. Of course this has been created in pockets of spare time in-between the real projects. I estimate it took ~52 x hrs, averaging 2 hrs a day.

Either way, I hope you enjoy my solo effort more than the actual Coke ad 😛

17/09/2025

Doesn't your campaign deserve the headliner?

15/09/2025

Three words and the truth.

This is the postcard that I sent out last week to some of my fav agencies and industry collaborators, reminding them that my pen is smoking and that I'm ready to body-surf across any brief that they throw into the mosh-pit.

But what inspired this postcard?

In the battle of the brands, I've always loved the way that a succinct and cleverly penned headline can transform complex information into a resonant 'power-chord' that cuts through the noise, inspiring not only action, but rousing a new generation of fans.

Thinking about the tension between headlining bands and the support acts that thanklessly strum in the face of apathetic, or worse, hostile crowds, an idea for a self-promo piece began to slowly build to a deafening crescendo, just like the impatient claps in a stadium filled with punters who want all killer, no filler!

Importantly, no AI was used in the creation of this artwork and I spent endless hours trawling op-shops for the perfect costume and accessories, growing my hair out, and air brushing the speed flames onto the giant novelty pen.*

Ironically, there isn't a headline so to speak. But the copy on the back of the postcard does the heavy lifting, underlining the fact that I am the 'headline act'.

So if you need unforgettable headlines, anthem worthy campaign ideas, and copy that goes to eleven, then let's grab a bevy backstage, snort a line of prose, and talk shop.

If you'd like to see a PDF of some of my heaviest hits then shoot me a DM or email me at [email protected]

"One, two, three, phwoar!"

*100% AI