Can AI Replace a Real Shoot? We Put It to the Test
- Pania Kirillina
- 21 hours ago
- 4 min read
A studio diary from the frontlines of AI-generated video production.
We didn’t set out to become full-time AI video testers. It just happened.
One of our long-time clients came to us with a request: a short product commercial, fully generated with AI. This request didn’t come as a surprise — we had already collaborated with several clients on product campaigns where it was mutually understood from the outset that AI would be used to produce the visuals. The results for most were shockingly consistent. With some smart prompting and, yes, still quite a lot of hand-tweaking, we got lifelike portraits, great styling, and crisp renders that passed for real photography.
So when our client asked, “Could we just try to do the whole thing with AI for our video campaign?” — well, the sentiment made sense. If this technology can deliver results close to a $200K video production — and you get to control every single detail — why wouldn’t you try it?
So we said yes to running some tests before deciding whether we wanted to go full speed into production. And that’s when the questions started.
Here was our prompt. Simple, clear, good vibes:
A red-haired woman in her late twenties, casually dressed in a blue t-shirt and white headphones, dances to the rhythm of music with a soft smile. She’s in a vibrant, modern art studio — full of color, light, and creative motion.
To our surprise (and slight relief), the platforms interpreted this prompt in refreshingly diverse ways. The results included women of various backgrounds, builds, and features — though, yes, conventionally attractive white women still made the most frequent appearance.
From there, the chaos began. Here’s what happened — platform by platform:
Hunyuan
Famous for its open-ended text-to-video flexibility, Hunyuan was our first stop.
Issues:
None of the outputs had models with open eyes. (Why are AI eyes always closed?)
Our heroine "danced" while pointy disco fingers passed through her headphones.
Her smile, while appropriately happy, was frozen and unchanging — an AI grimace with no expiration date.
Verdict:
Fun to watch, but unusable.
Rating: 4/10
Wan 2.1
Known for solid faces and decent fidelity, Wan gave us a fully-formed woman this time. She definitely looked more human.
Issues:
The smile was there, but her face was locked in a single emotional frame, and no prompt tweaks helped.
The closed eyes, why are they always closed?
Product tracking was unreliable: headphones frequently shifted in frame.
Verdict:
Beautiful stills. Half-decent 4-second motion. But don’t ask her to finish the dance.
Rating: 5.5/10
Veo 2 (Google)
Google’s cinematic darling. Promises “film-like” results. And it almost delivered.
Strengths:
Gorgeous lighting with consistent shadow logic — if the light is left, the shadows behave right.
The model had distinct, memorable features.
Her eyes were finally open. She even emoted a little.
Issues:
Small product elements (like headphone logos) went missing.
Her movements didn’t quite match human rhythm.
The cost was steep for what was essentially B-roll.
Verdict:
Closest to real filmmaking, but not without sacrifices.
Rating: 6.5/10
Hailuo
This one rendered fast — and looked the part.
Issues:
Every output played like it was underwater. Slow-mo vibes we didn’t ask for.
The model was visibly AI — the kind your brain instantly flags as fake.
Frame elements often melted into each other (at one point the cord split like a snake).
Her hands… defy language. Let’s just say they’re better left unmentioned.
Verdict:
For when you want to feel alive by comparison.
Rating: 5/10
01-Live
01-Live gave us something no one else did: a consistent avatar across generations. That’s the good news. The rest was unremarkable.
Issues:
The face looked emotionally divorced from the body.
Hair and headphones kept warping in and out of frame.
Each generation took 3–4× longer than other tools.
Verdict:
Great if your video concept is "Haunted animatronic".
Rating: 4.5/10
Runway 4
Surprisingly, the heroine looked almost identical to the one we got from 01-Live. (AI siblings, maybe?)
Positives:
Strong facial realism, blinking and gaze shifts included.
She even tried to mouth along with the music at one point.
Issues:
She had too many teeth. Just… too many.
Movement was still awkward and slightly stiff.
Verdict:
High-quality uncanny valley.
Rating: 6/10
Sora
The one with the hype. OpenAI’s Sora is supposed to be the future — and for once, the demo reel didn’t lie too much.
Positives:
The camera moved, too, creating real depth and parallax.
The model pulled off a couple of 360° spins. Completely unprompted. We’re not mad.
Her face made sense, her expressions tracked, and for once it didn’t feel like a haunted wax museum.
Lighting and skin tones were rich and cinematic — high contrast, clean, like someone had already done the color grade.
Issues:
A few shadow oddities and inconsistencies.
Occasional glitch in her movement and dancing.
Verdict:
This is a massive leap from where OpenAI video was just a few months ago. We’re hopeful.
Rating: 6/10
Kling 1.6
Kling gave us the widest variety of models and the most natural energy. It felt confident.
Wins:
Lively motion, good expression, strong character presence.
Little to no body distortion or morphing.
Backgrounds and atmosphere nailed the prompt.
Issues:
Close-ups struggled with product clarity — headphone details got mushy.
Some shots had minor glitching around movement edges.
Verdict:
The current frontrunner — but give it two more versions.
Rating: 7/10
Bonus: Luma (1/10)
Just to remind ourselves how far things have come, we threw the same prompt into Luma. Two months ago, this was the tool.
Results:
Hair looks and acts like an orange cloud.
Hands are the most AI thing ever — morphing into al dente spaghetti.
Everything is just melting.
Verdict:
A cautionary tale in motion.
Rating: 1/10
So… Is AI Video Ready Yet?
Not quite.
Kling, Runway, Sora and Veo are pushing boundaries. But even the best tools still fall short on product accuracy, consistency, and realism — especially for anything longer than two seconds.
That said, we do see potential for AI in specific use cases: short AI-driven loops (3–5 seconds), hero stills, or quick concepting passes. But for polished storytelling and campaign-level quality, traditional production is still the more effective path.
And while AI has dramatically improved in just a few months, we’re publishing this today… because next Tuesday, Kling might already be directing its own short film.
Curious to see more of what we do? Check out our work here.