I Tried AI for Video. Here's What I Got

I'm a developer who suddenly needed video

Duskroom is a thing I'm building: ambient sound rooms you leave open while you work or fall asleep. Eight scenes. Each one needs a short video loop underneath it - something calm enough to sit in your peripheral vision for twenty minutes without demanding anything.

And I have never made a video in my life. Not with a camera, not with After Effects, not with AI. So my starting position was: I know exactly what I want it to feel like, and absolutely nothing about how to get there. Knowing what you want and knowing how to get it are very different problems, and I only had the first one.

The rule I made up so I couldn't waste money

The AI video space right now is loud. Every YouTube video says a different tool is the best one, and half of them are affiliate links wearing a review costume. After a few hours of that I stopped trying to figure out the answer from the outside and made myself one rule instead:

Try it free. Pay only if the free result is already good enough to ship.

That's it. No comparison spreadsheet, no trial-and-error budget, no "let me subscribe for one month just to see." Every one of these tools has a free tier or a trial, and the trial is the review. Whatever a stranger on YouTube got out of it doesn't tell me what I'll get out of it, because they know how to prompt for video and I don't.

HappyHorse and Kling lost me in about an hour

I'll be honest about how shallow this test was: I didn't spend days with either of them. An hour or so each, maybe a handful of generations. Someone who actually knows video would probably get more out of them than I did, and I'd believe them.

But here's what killed it for me. Both tools kept dropping parts of my prompt. I'd describe a scene with three things in it - warm light, a slow camera drift, a specific mood - and get one of those three back. Then I'd rewrite it, simplify it, try again, and get a different one of the three.

The prompts weren't the problem - I'd already had Claude structure them, one idea per clause, nothing tangled. So when a model returns one of three requested things, that's the model's ceiling, not my writing. Combined with pricing that assumed I'd be here every day, I closed both tabs. Not a deep evaluation. Just a fast no from someone who couldn't afford a slow one.

The tool I liked most is the one I didn't buy

Higgsfield was genuinely the nicest thing I touched in this whole search. Dozens of models from different providers, all sitting behind one interface. For someone with no idea what he's doing, that's a real gift - I could try the same prompt across several models without opening five accounts and learning five UIs.

I wanted to pay for it. That's the honest part. I had the tab open with the pricing page for a while.

Then I asked myself a question I'd been avoiding: how many videos am I actually going to make this year? And the answer is eight. Maybe sixteen if I redo some. Duskroom isn't launched. It might become something real, it might quietly die on my drive, and I don't get to know that in advance. A monthly subscription doesn't care about that ambiguity. It just bills.

Higgsfield is priced for someone who opens it every morning. I'd be opening it once, then vanishing for months. The tool wasn't wrong. I was the wrong customer.

LTX 2.3 is free, and free is the honest price

LTX 2.3 22B costs nothing, no signup, I ran it through upsampler.com's free generator. It did the same thing HappyHorse and Kling did - took the first half of the prompt and dropped the rest. Every single time, very consistently.

That consistency is the useful part. Same structured prompt, same failure, in exactly the same place. When your input is controlled, the difference between five tools is the tools. That's the whole reason I bother structuring prompts before sending them anywhere: it turns a vague "this one felt worse" into something I can actually point at. LTX has a small context window. It doesn't fail gracefully. It just takes what it can hold and moves on.

I wouldn't put its output anywhere near a product. I'd still tell anyone to burn an evening on it before spending money elsewhere, because it costs nothing to find out what "not enough context" looks like.

The one thing I already knew before I started

I came into this green about video, but not about talking to models. I've been building with AI long enough to know that the prompt is most of the outcome, so I never typed raw into a video tool. Not once, not on the first trial.

The process from day one: find reference images and video for the mood I'm chasing, hand those plus my description to Claude, and have it return a structured prompt - one idea per clause, nothing tangled, phrased for a model rather than for a person.

The reason is boring and financial: a weak prompt burns exactly the same generation as a strong one. On a small quota a retry isn't free, it's a scene you don't get to make this month. Same instinct as keeping Claude Code and Obsidian in sync for my notes - do the structuring once, up front, so nothing downstream has to guess.

Using AI to write for AI still feels slightly ridiculous. It also keeps working, so I've stopped arguing with it.

What $5 got me, and where I'm stuck

I landed on Gemini Omni through a Google AI Plus promo: $5 for the first three months. I already liked what Gemini did with images, so it wasn't a blind bet, and enough forum threads confirmed the price was real and not a bait number.

The output is 7.5 out of 10 for what I need. Short ambient loops, not cinema. It doesn't make me gasp. It goes into the app with almost no cleanup, which for my purposes is the whole point.

The missing 2.5 is worth naming, though. The clips are calm and they loop cleanly, which was the actual requirement. They're also slightly more generic than what was in my head. The mood transfers. The specific detail doesn't, not fully. That gap is the honest cost of not knowing how to art-direct a model, and no budget closes it. Only practice does.

Then two things I didn't see coming.

The quota is small. Much smaller than I assumed when I saw the price. Three of Duskroom's eight scenes exist right now, and that's not because I stopped working - it's because I ran out. If you're pricing this out for a real project, the $5 buys you access, not volume.

And the watermark is still there. Every video, paid plan or not. Gemini shipped a "Media Watermark" toggle in settings built specifically to turn this off. I turned it off. Nothing changed. That toggle only arrived mid-August 2026 and is reportedly rolling out gradually, not to every plan or region at once, so I'm treating it as a release still in motion rather than a lie. For now every clip goes through a separate watermark remover before it touches Duskroom.

Five scenes left, a quota that says not this month, and a toggle I keep checking like it might have quietly started working.