THE SECRET TO BETTER AI VIDEO MIGHT BE LESS AI

How I’m working with AI video in the creative advertising industry (and not destroying my soul)

AI video is everywhere. It’s starting to appear in TV, social media and advertising.

That half-contrasty, over-sharpened, under-detailed, logic-less slop could have been the kiss of death for agencies who dared to make these a year ago, but not so much these days.

I’ve been experimenting with AI images and video generation for years. At first it felt like a toy trying to imitate what I do. Then suddenly social media filled up with cinematic, blockbuster-level AI films where the characters behave normally, and the physics feel believable.

Epic, cinematic, blockbuster-level videos that look like they were made by a production team with millions in funding. Characters act normally, simulated physics feel realistic.

So how are these people making magicwhile everything I try to fit into my workflow systematically dismantles my will to live?

Adverts aren’t made by one personin their studio apartment banging out AIgenerated shots until something looks good.In this industry we have deadlines, iterations, collaboration, branding, products, iterations, dos, don’ts and inevitably surprise client stakeholders who pose their ‘what-if?’, ‘also’, and ‘how about’ iterations at the last minute.

So you generate a character in a fluffy orange mohair cape and Nike trainers, does a front flip over a charging hippopotamus, kicks a football into a goal, landing on a quad bike and disappearing into a cloud of dust as the ground turns to lava.

Nice.

Then the client asks for less dust.

The hippo needs to be scarier.

Maybe make it a rhino.

Can we put “Just Do It” on the quad bike wheels?

Who knows what you’ll get if you click Generate.

Traditional VFX and compositing arebuilt around modularity. Live-action plates, backgrounds, dust layers, 3D rigs, tracked assets and colour grading can all be adjusted independently.

With AI, you often have a prompt, a reference and a result.

No layers. No data flow. Everything baked in.AI works through directed randomness.

That makes it incredible for ideation. A prompt can produce a strong first look or endless variations incredibly quickly.

But professional creative work is about intent.

Did we really choose the talent, action, costume, props, lighting, environment and lensing? Or did the model interpret a sentence and make half those choices for us?

This is where references becomeinteresting.

With references, our prompt for thatspecific element is immediately more tangible.Show the model the subject wearing the costume, holding the product and doing the action. Photograph the environment from multiple angles, create a 360º stitch, or build it in 3D so you can recompose and retexture it.

Move the product 20 pixels right.

Make the rock 20% smaller.

With all this, we’ve got a near-perfect reference for the generative model to use.

Like many others, I had been initially reluctant to integrate generative AI into my workflow.

My biggest objection to generative AI has never only been the wider concerns around environmental cost, labour displacement or creative over-reliance.

It’s been control.

Professional work depends on precision and intent. Raw generation still puts too much of the outcome in the hands of chance.

For me, AI becomes much more usefulwhen it stops trying to replace the production process.

When grounded by live-action, 3D andcompositing, it becomes less of a sealed black box and more of a responsive creative tool.

One that can adapt, iterate and still leave room for the unexpected.