The Making of a Video Trailer

Making this video turned out to be quite fun – and educational, in AI.

See the video first, if you want, here –

Step 1 – The Storyboarding.
This is manual. I can’t ask AI to write a script yet; I tried, with fairly detailed prompting, but the results were average at best, and very, very clichéd. I had to site and visualize how I wanted this to flow, break it into scenes, and then generate each scene independently.

Step 2 – The Video Shots
This was quite interesting. I knew I wanted cinematic/realistic, short clips, so focused on finding platforms accordingly.
Imagine shorts – tried to rewrite the prompts into something else that would’ve deviated too much from the original, rejected.
Hailuo AI – very good. The movements are smooth, realistic, backgrounds detailed, and it has a nice control over camera movement (limited to what’s possible within the 6s tests). I’m seriously considering this as a paid version for more stuff. The demo lets me generate 4 tests, 6s each, watermarked, at 768p resolution.
Fotor – Never got to test – too few credits for video generation from text, though it looks pretty capable given it’s using Veo3.
GenApe – The worst of the lot, despite glowing reviews (maybe my prompts weren’t good enough, but still) – blocky, stiff, primitive-looking glitchy output, and a render time of close to an hour. I gave up pretty fast here.
Google AI Studio – also pretty good. 8s clips, 720p res, minimal prompt reworking, clean and near-perfect output
OpenArt – also quite good, and kind of a cheatcode since it also uses the Hailuo engine, so I essentially used it twice. It has options for automatic audio and lip sync (though I didn’t test these) but not camera movement. However, it does have access to use multiple other generators, so this may be a better bet for testing multiple genAI video tech and quality. Output was 6s, 768p, watermarked.

What I found AI video good at – very useful when you –

1. Focus on clear, unambiguous, foreground events.
2. Use stereotypical visuals, they come out well – think what kinds of videos exist in huge numbers and variations, to provide sufficient training data.
3. Be very, very clear, detailed and visual in your prompt descriptions, and you get something close.
4. Be prepared to rewrite and expand the prompts exhaustively. You’ll never get what you want on the first try.

The hardest areas, or where AI struggled –

1. Video-in-video. The initial shot which I thought was pretty basic, ran through dozens of failures – the movements were stiff and jerky, or comically exaggerated, while the background screens displayed almost uniformly off-the-mark, basic, stereotypical content, or grossly primitive-looking 4 generations ago output (eg. a drone shot of a beach showed up as a drone ON a beach. with giant cartwheels.), or got stuck with rendering still images.
   In hindsight, this makes sense – In requesting a screen wall, I was asking for not one, but 10 separate videos. Obviously allocated resources will render 10% of quality.
   Even something as basic as the number of screens (3X3) never quite worked out – I guess video AI can’t handle the simultaneous display structuring and resource allocation within that structure together.
2. The other area, surprisingly, should have been easy, but wasn’t – blank voids and empty spaces. They felt noisy, randomly shadowed, flickering, or flat and unrealistic. I guess there isn’t sufficient training data on zero-content video.
3. And don’t even get me started on text-in-video. GenAI in video simply cannot render text. Explicit instructions, clearly defined words, sample image guides, nope. A gross mishmash of misspelled letters glitching everywhere. The big red question mark at 0:28 is the fallback option after I had to give up the original vision.
4. Backgrounds. Blurred or static ones are fine, well-rendered, but add movement, and it goes berserk. One butler becomes an army, marching back and forth from all directions. A gesture of surprise becomes a comical melodrama. And background elements don’t interact (I guess this is the rendering capacity limits for vid-in-vid issue again)
5. Replication. Throw a paper in the air, it becomes a veritable waterfall. A calculator on a desk means a Radio Shack-esque display counter of calculators. (I know I’m dating myself but what current retailer sells calculators?). Spit becomes a hosepipe. Subtlety, thy name is not AI.

Step 3 – The Audio

While not as exhaustively tested as video (I just tried 2, Veed and ElevenLabs) I found audio almost universally behind video in quality of output. I guess while video allows sufficient detail to come through that it overwrites the judgmental ability of the eye, audio is still very clearly in Uncanny Valley – speech sounds stilted, stiff, robotic.

And who knew a Hollywood action trailer announcement voice would be so hard to find?
Elevenlabs delivered, but some glitch prevented account creation, so I couldn’t keep variants. Veed didn’t have the voice I wanted, but some interesting ambient sounds and effects capability that could be useful in future.

Step 4 – The Production

For an amateur not used to video editing, this was a revelation. Not at how good video editing software has gotten, because it hasn’t. Tested 3 different editing packages, and they were universally awful. Complex, glitchy, and somehow still very limited in capability. In fact, they were so bad, even compared to good old Windows Movie Maker (which ran on Win 3.1!!) they were effectively unusable.
Instead, (surprise!) it was Canva to the rescue. It’s an excellent solution if you just need to quickly stitch together a couple of clips and an audio track – a simple, familiar interface, all the capabilities of WMM without the RAM-burning, CPU-overheating struggle, and nice transitions, effects, cropping, and speed control. Plus, it lets you add still images and modify in the same video stream, so yay for intro-outro screens.

That’s it for now – Suggestions, tips, and ideas welcome, if you’ve had similar experiences. Tell me what worked for you, what you found useful – would love to test it out.

Links to the platforms used –

Imagine Shorts – https://www.shorts.imagine.art/dashboard
Hailuo AI – https://hailuoai.video/create
Fotor – https://www.fotor.com/apps/ai-video-generator/
GenApe – https://app.genape.ai/ai-video-generator
Google AI Studio – https://aistudio.google.com/
OpenArt – https://openart.ai/video/t2v/MiniMax-Hailuo-02
ElevenLabs – https://elevenlabs.io/
Veed – https://veed.io
Canva – https://canva.com