FilmSkool — a free film school
The AI filmmaking primer
The AI Filmmaking department is 145 verified lessons and not one of them answers "what is a prompt". This is the part nobody teaches, written here rather than borrowed.
18 entries · Original writing
The eight words nobody defines
Every tutorial in this department assumes you already know these. None of them are complicated, and not knowing them is the only thing standing between you and the rest of the wing.
- Model — A very large file of learned statistical patterns, which you hand an input and get an output back from.
- LLM — A Large Language Model — a model trained to predict the next piece of text, which turns out to be enough to hold a conversation.
- Prompt — The input you hand a model — usually text, often more than text.
- Token — The chunk a model actually reads — roughly three-quarters of a word — and the unit almost everything is billed and limited in.
- Context window — How much the model can hold in mind at once. Past the edge, it has not forgotten — it never saw it.
- Diffusion — A model trained to remove noise, run backwards from pure static until a picture is left behind.
- Seed — The number that decides the starting static — and therefore the one control that makes a generation repeatable.
- Hallucination — Because it is producing plausible output, and plausible is all it was ever trained to produce.
- MCP — Model Context Protocol — a standard way to plug a model into your actual tools and files, so it can do things rather than only describe them.
- Agent — A model put in a loop with tools, so it can take several steps toward a goal instead of answering once.
Image models and video models are different animals
The single most useful thing a beginner can understand about this field, and the reason almost every good AI film is made stills-first.
- Image model — Producing one frame at very high quality, with a great deal of control, very cheaply and very fast.
- Video model — Good at making a handful of seconds move convincingly. Bad at length, continuity, and doing what it is told.
- Stills-first workflow — Because you get all of the image model’s control and quality, and only ask the video model for the one thing it is actually good at.
- Image-to-video — Hand the video model your frame as the starting point, plus a short description of the motion you want and nothing else.
The problems nobody has solved yet
Knowing where the ceiling is saves more time than any tutorial. These are the things that will break your film, and the ways people work around them.
- Consistency — Because nothing carries between generations by default — each one starts from fresh randomness with only your prompt to go on.
- LoRA / fine-tune — Train a small add-on file on a handful of images, then load it alongside the base model.
- Sound — Because almost nobody does the sound, and sound is half of what makes an image feel real.
- Disclosure and rights — Less than the tools imply, and the honest answer varies by country, by platform and by what the model was trained on.
Where this is taught in the catalogue
This is the plain index, shown because JavaScript is off. Every lesson here is a free video published by the channel credited on it and plays through that channel's own embed — FilmSkool hosts nothing.