Level 1 / Compositing
AI for VFXHybrid Workflows
Generate the element with AI. Integrate it like a compositor.
Read the first lesson without an account. Nothing to buy, nothing to install.
this is the entire trick
- 36
- Lessons
- 3 h
- Reading
- 20
- Things to drag
- 25 of 36
- Free to run
What this is
You work on one supplied plate for the whole course. Every lesson advances the same shot, so nothing here is a disconnected tool demo, and you finish with one piece of work rather than thirty exercises.
The generation half is the easy half, and it is the half everybody else teaches. Most of this course is the other half: colour space, alpha, edges, defocus, light wrap, contact shadow and grain. That is the part that decides whether anyone believes the shot.
Every integration lesson ends in a measurement you read off your own image. The point is not the settings, which change when a model ships. The point is that you can check your own work forever afterwards without us.
Who it is for
Artists new to AI who can already navigate a node graph or a layer stack: junior compositors, motion designers, students and VFX adjacent generalists. Also for AI native creators who can generate anything and cannot make it sit in a shot.
What you need
A browser. No local software and no GPU. You should know what a layer, a frame and a mask are. No colour science, prior AI generation or scripting assumed.
Who it is not for
People who want to make a whole AI short film. That is a different course and honestly a different product.
The whole thing
9 modules · 36 lessons · 108 questions
You can get a long way clicking generate and hoping. You cannot get good that way. This module is the seven things that turn "why did it do that" into "of course it did that", and every one of them stays true when the models change.
It is not a folder of pictures
The model kept the patterns and threw the pictures away. That is why it can make something new, and why it cannot make your shot.
It was taught by wrecking pictures on purpose
Destroying a picture is easy and you always know the answer. So they destroyed a lot of pictures, and trained a machine to take one step back.
Making something new out of static
Hand a cleaner some static and it will clean it into something. What that something is, is decided by the static.
Your words are a nudge, not a command
The prompt leans on every cleaning pass. It does not instruct anything, and leaning harder eventually breaks the picture.
The static is the only thing that changed
Same words, different static, different picture. Lock the static and you get a laboratory instead of a slot machine.
It works small, then enlarges
Almost all the thinking happens on a small version. Small things live in few pixels, so small things are where it guesses hardest.
Moving pictures, and the three ways to steer them
A clip is denoised as one lump, not frame by frame. Everything good and everything bad about video follows from that.
Why it is built this way
Most online courses are abandoned. Median completion across large open courses sits around twelve percent, and the reasons are consistent enough to design against: passive video, courses that run too long, and material that assumes knowledge the student does not have. Every structural decision below is an answer to one of those.
You drag things, you do not watch things
Across 225 studies, active learning raised exam performance by roughly half a standard deviation and students in traditional lectures failed at about one and a half times the rate. So every lesson has something you operate yourself.
Every lesson ends in questions
Being asked to recall something beats reading it again, and the gap widens over time. Rereading feels more productive, which is exactly the trap. Only your first answer scores, so the number stays honest.
Nothing assumes you were in the room
The Foundations module teaches how these models actually work before anything asks you to use one. Skipping that is the most common reason people bounce off technical material.
Short lessons, and no locked doors
Longer courses complete worse, so most lessons run four to seven minutes. Nothing is gated behind a subscription and nothing is gated behind a credit balance: 5 of the 9 modules run entirely in free nodes.
The evidence, if you want to check it
Active learning increases student performance in science, engineering and mathematics
Freeman and others, PNAS 2014
The 225-study meta-analysis. Exam performance up by roughly half a standard deviation, and students in traditional lectures failed at about one and a half times the rate. It is why this course makes you drag things rather than watch someone else drag them.
Retrieval-Based Learning: A Decade of Progress
Jeffrey Karpicke
Why being asked beats reading it again, even though reading it again feels far more productive. That feeling has a name, the illusion of competence, and it is the reason every lesson here ends in questions.
The Past, Present and Future of the Cognitive Theory of Multimedia Learning
Mayer, Educational Psychology Review 2023
Two hundred-odd experiments on how to lay out a lesson: words with pictures beat words alone, and anything decorative actively costs you. The spacing and figure placement in this course come from here.
Uncovering MOOC Completion: a comparative study of completion rates
Open Praxis, 2024
Median completion around 12.6 percent, and lower for longer courses. Read as a list of things not to do: passive video, six-week blocks, and assuming knowledge the student does not have.
Everything worth reading
Papers, standards and the occasional beautiful explainer. No course sellers, nothing behind a paywall, and every link checked. If one of them dies, our build fails.
How AI Image Generators Work (Stable Diffusion / Dall-E)
Computerphile, with Dr Mike Pound
The friendliest explanation of denoising that exists. If one thing in this module has not clicked yet, watch this before you read it again.
What are Diffusion Models?
Lilian Weng
The standard reference write-up. It gets mathematical fast, so treat the first three sections as the useful part and stop when your eyes glaze. Nothing later in this course depends on the equations.
Denoising Diffusion Probabilistic Models
Ho, Jain and Abbeel, 2020
The paper that made this whole industry work. You do not need to read it. Skim the pictures once so you know what a real one looks like, because you will hear it cited for the rest of your career.
High-Resolution Image Synthesis with Latent Diffusion Models
Rombach and others, 2021
This is the one that explains why your fine detail goes soft: the work happens on a small compressed version of the picture. Read the abstract and look at figure 3.
Classifier-Free Diffusion Guidance
Ho and Salimans, 2022
Where the guidance slider in every generator you have ever used comes from. The abstract alone tells you why cranking it does not get you a more obedient picture, it gets you a worse one.
Video Diffusion Models
Ho and others, 2022
The jump from one picture to many. Worth a skim purely for the intuition that a clip is denoised as one lump, not frame by frame.
Stable Video Diffusion
Blattmann and others, 2023
Useful for one thing: seeing how much of a video model is decided by what it was fed and how it was curated, rather than by clever architecture.
Alpha Compositing
Bartosz Ciechanowski
Genuinely beautiful and fully interactive. Drag the sliders. Twenty minutes here will teach you more about premultiplication than a year of guessing.
Compositing Digital Images
Porter and Duff, SIGGRAPH 1984
Eight pages from 1984 that still describe exactly what your Merge node does tonight. Historically lovely and surprisingly readable.
Image Compositing (CS148 lecture notes)
Stanford Computer Graphics Laboratory
Slides, so it reads quickly. Good for seeing the same ideas drawn a second way when a diagram here has not landed.
Cinematic Color
Visual Effects Society technical committee
The industry write-up on how colour is actually handled on a show. Dense, but it is the document your future supervisor half remembers.
Colour FAQ
Charles Poynton
Old-school web page, no design at all, and still the clearest answer to "what is gamma really". Keep it in a tab.
ACES documentation
Academy of Motion Picture Arts and Sciences
What a facility means when they say "we are an ACES show". Read the overview only, unless you enjoy this sort of thing, in which case enjoy.
OpenColorIO
Academy Software Foundation
The colour plumbing under most of the software you will ever touch. Worth knowing the name so you understand what a config file is doing.
OpenEXR technical introduction
Academy Software Foundation
Why the format everyone delivers in holds values above 1.0, and what a channel actually is. Short, and it demystifies a lot.
Recommendation BT.709
International Telecommunication Union
The actual document behind the words on your delivery spec. You will never read it end to end. Open it once so it stops being a magic phrase.
AV1 at Scale: Film Grain Synthesis
Netflix Technology Blog
A streaming service caring enough about grain to strip it, describe it and rebuild it on your television. The best possible argument that grain is not noise.
VFX Reference Platform
Academy Software Foundation
One table that tells you which library versions a facility is standardised on this year. Useful context for why studios move slowly.
fxguide
fxguide
Long-form breakdowns of how real shows were actually made. Not a course, not a tutorial mill: read it the way you would read a trade paper.