The real distance of AI-generated video: a 15-second ad, NT$42 and a full day

On 5 August 2026, I spent a day building a 15-second vertical ad from scratch. Every frame was AI-generated. Raw material cost: NT$42.

This piece isn't about how amazing AI is. It's about what that NT$42 bought — andwhat it couldn't buy. Because the second half is what actually decides whether you should use it.

Why run this experiment

I wanted a Reels ad for my own LINE bot, "94eat". The premise is simple: an office worker can't decide what to eat at lunch, uses the service, walks out, and ends up with a bowl of pasta.

The traditional route: cast talent, book a location, shoot for a day, edit for a week. Even at the bare minimum, you're into five figures.

So the question is blunt — how far can AI video generation actually go right now?

The tool I picked was LibTV (an AI video platform from China's LiblibAI). It ships an official CLI, so you can drive it with commands and wire it into a pipeline. Worth noting: it's a China-based platform — registration needs a mainland phone number, payment goes through Alipay, andthe servers are in Beijing, so a direct connection from Taiwan doesn't work. I ended up routing API traffic through my own cloud host to get in. That alone tells you part of the cost never shows up on the invoice.

The actual price list

The platform bills in "credits". Unit prices I measured myself:

Action	Credits	Time
Script → shot list	6	73 sec
One storyboard frame	1	34 sec
5-second clip (720p)	40	2–10 min
15-second clip (720p, single take)	120	8 min
One TTS voiceover track	~31	23 sec

The pattern is clean:video is 8 credits per second, regardless of which model you pick — 480p halves it, 1080p doubles it. Only the flagship model is priced separately, at 3.35× the rate.

I bought the ¥59/month plan (about NT$254) with 1,500 credits a month, which works out to roughly NT$0.17 per credit。

So:

• The raw material for a 15-second ad costsNT$26 if it goes smoothly, NT$42 including failed attempts
• The full day of experimenting burned 546 credits total,NT$92

⚠️ One thing that's easy to misjudge:subscription credits reset every 31 days and don't roll over. The annual plan looks cheaper, but you can't bank credits and spend them all in the month a project blows up. That matters a lot if you work on client projects.

The first real problem: not image quality, but consistency

Image quality stopped being the issue a while ago. The Taipei arcade I generated had scooters, red shop signs, harsh midday light — unmistakably Taiwan. The pasta restaurant was full, steam rising, background diners softly out of focus.

The real difficulty is this:the lead in all three shots has to be the same person.

I tried three approaches, with wildly different results:

Approach	Scene	Character
Do nothing	✅ Scenes look great	❌ Three shots, three different people
Feed the character reference sheet in as a "reference image"	❌ Scenes destroyed	✅ Perfectly consistent
Write the character's appearance into the prompt as text	✅	✅

The middle row is the counterintuitive one. I assumed "give it a photo of the character" would be the most direct fix — instead, the model copiedthe reference image's background and composition too. Office, street, pasta shop: all three scenes turned into the grey studio backdrop from the reference sheet, leaving a single plate of pasta stranded on a wooden table against grey.

The reason isn't hard to see: image-to-image takes the whole image as its starting point. It can't copy the person without copying the background.

What finally worked was the crudest method: no image at all — just stuff "shoulder-length black hair, grey short-sleeve top, beige wide-leg trousers, dark backpack" as text into every single shot prompt. All four frames matched exactly on hair and wardrobe, and all three scenes survived.

The lesson here generalises to every generative tool: more reference material doesn't necessarily mean more control, because you can't specify "only reference this part". Sometimes a text description is more precise than an example.

The second real problem: the less camera movement, the more stable

In my original shot list, the second shot was "camera tracks alongside her". When the clip came back, the client's first reaction was "it looks like it's dropping frames here".

I measured it — all 150 frames were distinct, no duplicates anywhere. Technically completely normal.

The problem was somewhere else. In a tracking shot, her position in frame stays fixed while the background slides horizontally — but in the AI-generated version,the frequency of her leg movement didn't match the background's speed. The brain catches that mismatch and reads it as stutter.

I tried switching to a locked-off camera, which came out worse. I wrote "keep the background static" in the prompt, meaning to lock the camera — the model read it as locking the background, freezing most of the passers-by into statues while a few kept moving. Half-frozen is stranger than either fully moving or fully still.

What finally worked waschanging direction: have her walk toward the camera.

With movement along the depth axis, displacement is carried mostly by the figure's changing size, so the legs and background don't need to sync tightly. It worked first try.

All three shots ended uplocked off. And the one that was most stable from the very first generation was the only one that never had camera movement in the first place.

You can take this one straight to the bank:when generating video with AI, default to a locked-off camera unless there's a reason it has to move.

The third problem has no fix: no seed, only rerolls

The first two problems can be solved with better prompts. This one can't.

In traditional editing, you say "hold this half a second longer", "that expression is too big, dial it back". That'sadjustment。

Here, adjustment doesn't exist. There's onlyrolling the dice again. I revised the second shot's prompt three times, and each time I got an entirely new clip — not the previous one fixed, but a different one regenerated from scratch. These models don't expose a fixed random seed, so "same thing, one change" is simply not available.

This limit is more serious than it sounds. It means:

• You can't fine-tune — you either accept it or start over
• Every restart costs the full price and the full wait
• You never know whether the next one will be worse than this one

The first generation of my third shot had her staring straight down the lens the whole time, expression overplayed, with the pasta in her mouth vanishing between frames. The second pass — "eyes on the food, not the camera, restrained expression" — nailed it. But that's partly luck: run the same prompt again and you'll get something different.

One thing AI can't generate: real product footage

My original script called for a map, finding a restaurant, actual LINE conversation screens.

Thosecan't be AI-generated. Any generated LINE interface is inevitably fake — button placement wrong, cards wrong, typeface wrong. And using a fake interface to advertise a real product isn't a technical problem, it's an integrity problem.

So the whole structure of the script had to change: every shot with a phone in it hasthe screen facing away from camera, showing only gestures and expressions. The ad ends up selling "that walk after the decision is made feels light" rather than "look how good this interface is".

In hindsight, the constraint made the ad better. But that was an accident, not a design decision.

Any product ad that needs to show a real interface — you have to shoot that part yourself. There's no way around it.

So what is it, really

After a full day, my conclusion is this:

it's a materials factory, not a finished-film pipeline.

What genuinely gets automated is substantial:

• Drop in a script and it breaks it into a shot list automatically, with shot size, lighting, and camera notes. The quality is surprisingly high — it picks up setups and payoffs (my script mentioned "an old burn scar on the back of her left hand", and it wrote that in specifically on the hand close-up)
• Scene generation needs almost no fixing
• Character consistency is solved
• 15 seconds in a single take, zero drift

But if you need to deliver a finished film, post-production hours barely shrink:

• The voiceover came out at 10.6 seconds, the visuals at 15.1 —audio and picture are generated independently, with no sync mechanism whatsoever
• End cards you type yourself (asking AI to generate Chinese text produces garbled letterforms)
• Music you score yourself
• Every shot comes out a tidy 5 seconds, but real editing is "3.2 seconds here, 6.8 seconds there"
• To have a character speak on-camera in sync, you need the audio file before you can generate the visuals — the whole order reverses

Who should use it, and how

At this stage I see three uses that pay off:

1. Situational B-roll. Those "people living their lives" cutaways in an ad — walking, eating, zoning out, checking a phone. Cut alongside your own live footage, the cost drop is very noticeable.

2. Previsualisation. Generate composition and pacing before the shoot — faster than a hand-drawn storyboard, more concrete than describing it out loud. At NT$0.17 a frame, you can "shoot" the whole thing once before deciding whether to shoot it for real.

3. Short-form social assets. Where a bit of roughness is acceptable, ship it straight out.

What doesn't pay off is "a finished film delivered to a client". That distance is still there.

Finally

NT$42 and a full day. The NT$42 bought raw material; the day went intoworking out where the boundaries are。

I think that's a good trade. Once you've mapped the boundaries, the next film won't take a full day — and I know exactly what it can and can't do, so I won't promise a client something I can't deliver.

That's probably the most practical way to use any AI tool: pay the cost once to measure the ceiling, then work inside a range you actually know.