Part 1

Where to generate

01 What these platforms are

Most of the models that make video do not sell to you directly. Some have no user interface at all, some are closed to certain countries, and some change their access rules every few months.

So people do not go to the model. They go to a platform that has already settled all of that: one account, one balance, a list of models inside. You pick a model from a dropdown the same way you pick a font, and you can switch between them quickly and easily.

02 The platforms

They are not ranked. They take your money in different ways and they let you in on different terms, and which of them suits you is for you to decide.

Higgsfield

A web app made for people who produce short video. You open it in a browser, pick a model, write, generate. Video models, image models, upscaling and the tools around generation all sit in one place.

You pay a monthly subscription and it gives you credits. Every generation spends credits, and the amount depends on which model you use, how long the clip is and what resolution you ask for. A cheap model and an expensive one on the same plan can be ten times apart.

Worth knowing before you pay: unused credits do not carry over to the next month. Whatever you have not spent disappears when the cycle ends. So the real price of a month is not the price of the plan, it is the price of the plan divided by what you actually generated.

Go to the site

fal

A site with access to all the models, and, most important, you can pay for each generation separately rather than having to buy a package.

You put money in first and it is drawn down as you use it. No subscription, no monthly reset. Most video models charge per second of finished video, some charge a flat amount per clip. You pay for the results you got.

There is no monthly reset, so a quiet month does not wipe your balance. The money does not sit there forever either: purchased credits expire a year after purchase, and any free or promotional credits expire after ninety days.

Go to the site

Syntx

A web app and a Telegram bot on one account, with a long list of models inside: text, images, video and voice. You buy a plan, it comes with tokens, generations spend tokens.

For a lot of people the reason it exists is access. It works from places where the model's own site does not, and it accepts payment methods the big platforms reject, cryptocurrency included. If you have run into "not available in your country", or had a card declined with no explanation, that is the problem it solves.

Tokens do not expire. They stay on the account while the subscription is active, and a new subscription adds to what is left rather than wiping it. A quiet month costs you the price of the plan, but not your unspent balance.

There is a small free quota to look around with, five tokens. Enough to see the interface and not much else.

Go to the site

Figma Weave

Also a web app, built as a canvas with nodes. You assemble a chain: generate an image, crop it, hand it to a video model, raise the resolution. The chain is saved and runs again with different inputs. Inside there are image models, video, 3D and ordinary editing tools.

The subscription gives a monthly pool of credits, and different models eat different numbers of them. You can buy more credits on top of the plan.

On the free plan there are no video models at all, only images; video opens up on the first paid plan. Unused credits carry over on some plans and not others: on the lowest one they burn, on the higher ones they build up, but no further than three monthly allowances. Credits bought on top last a year.

Go to the site

Magnific

Everything at once in one package: image, video, music and voice generation, editors, upscalers and a stock library of hundreds of millions of ready files. Inside is nearly the whole known list of video models.

You can pay monthly or for a year up front, and paying for the year is cheaper by about a quarter. Credits then come as a yearly pool: they last a year and are not zeroed at the end of the month. You spend them whenever you like across the year.

On the higher plans some models are declared unlimited. Read that literally: the unlimited part covers a specific list, mostly images, and video always eats credits.

Go to the site

Krea

A web app with a canvas where images, video, 3D, upscaling and training a model on your own pictures all sit in one place.

Go to the site

ElevenLabs

Text to speech, voice cloning, sound effects, music, translation and redubbing a finished video into another language. The same place also does image generation and editing and the whole working list of video models, Veo, Wan, Kling, Seedance.

A subscription with one shared pool of credits across all the tools at once. The same pool goes on voice, on effects, on music and on dubbing, so a heavy redub eats what you were saving for voice work.

The free plan does not give you a commercial licence. You can generate a voice on it; you cannot put that voice into a clip you earn money from. Commercial rights open up on the first paid plan, and that plan is cheap.

The strong side is still sound, and it solves a separate job: in a finished clip the audio is rarely taken whole from the video model. A line, an effect or the music is usually finished separately, and this is where that is done.

Go to the site

There are far more of them

There are dozens of these platforms.

So check everything written here about specific platforms yourself before you pay: plans, limits and model lists change constantly.

03 Subscription or per generation

There will be no prices in this course (almost), and that is deliberate.

On every platform the cost changes constantly. Discounts start and end, new models arrive and pull the price down, old ones get more expensive, a model moves from one plan to another, and the plan itself gets renamed.

The checking is on you, and it takes a minute. Open the pricing page and look not at the price of the plan but at the cost of one generation on the model you came there for. Or hand those pages to your chatbot and ask it to count: how many attempts on the model you need does each platform give you for the same money.

Part 2

Sharp against polished

There is one catch. The sloppiness was the point. Sora was, in my opinion, the sharpest model of them all. By sharp I mean the speed at which things move across the screen, and that speed multiplies how real the picture reads.

The scenarios are often nonsense: cats firing rifles, cats walking out of shops with stolen bottles, groceries, game consoles. The picture still reads as the most believable of the lot, because it looks like something filmed on a phone. The blur, the softness, the 720p: that is where the charm comes from.

In my experience there is currently no model that reproduces the Sora result at 100%. Kling, Seedance, Google Omni all lean cinematic. Seedance will hand you a frame that looks like a Hollywood film. What it will not hand you is footage that looks like a passer by pulled out a phone because two cats were fighting on roller skates. It will produce the clip, and the clip will be realistic in its own way, but the sharpness and the aliveness will not be there. You get a polished picture, and instead of a hard fight with hissing you get two cats gently patting each other.

Seedance handles people better, but that is not what we are here for.

The model does not read your mind

Do not fall into the trap of thinking the model reads your mind. It does not.

Part 3

Ideas

As I said earlier, the idea is the thing you are supposed to have already. So far no AI is going to come up with something interesting and new for you. At best it will be some average, generic idea. To get an AI thinking and producing ideas you actually like, you have to run one long conversation, a project, where you go through hundreds of ideas with it and explain what lands and what does not. After many iterations of that, the AI MAYBE starts handing you not 100% garbage, but, say, only 90% garbage.

That is how it works, and that is how I come up with ideas. Sometimes the whole idea arrives on its own. Sometimes, when there is no inspiration, I go to the AI for help. But in almost 300 clips on the account there has never once been a case where the AI came up with the idea in full. Never.

At best it gives me 20 or 30 ideas pointing in different directions, and SOMETIMES one or two of them are not complete garbage but something you can work with. That is an impulse to invent the material, not the finished material.

Here is how that looks. One of the AI's ideas, word for word:

Wet cement. A pavement. A ginger road worker smooths the last corner of a fresh slab and steps back with the float in his paws. A passer-by crouches down, presses his whole palm into the cement, adds his initials with a finger and walks off, shaking his hand clean. Ending: a micro freeze frame and a hiss.

That is the skeleton of an idea. In the first part the AI came up with a decent setup, but because it does not actually understand what is funny and what is not, it then writes complete nonsense: for some reason a random human walks up to a cat who is smoothing cement and, for no reason on earth, presses his palm into it. Why would any human who is not an idiot do that?

So the idea is half ready. I can see it will be funny if it is done not by a human but by another cat, because cats love ruining things, knocking glasses off tables, dirtying floors and so on. That is logical and clear to any person on earth speaking any language. Ha, the silly cat walked across the cement and ruined the other cat's work. You get it in a second, it is obvious and it is funny, because it is a cat, an animal, it has no idea what cement is and honestly walks over it without understanding that it is ruining it. The AI's version, where a human does it, is madness, and the only emotion it produces is confusion: what is this rubbish, why did the human do that, is he stupid?

And here is the clip that came out of it:

Over many iterations I arrived at one prompt that I use to find ideas: the formula of a funny scene, four filters every idea goes through and a kill list of what never works. It is in the full course.