AI Prompt Engineering Course Picks for Image and Video

If you want an ai prompt engineering course because you're trying to get better at generating images or video, most of page one is going to waste your afternoon. The big platforms own these results and their courses are text-prompting syllabi, taught with chat examples, ending before the part you came for.

The direct answer is that generation-focused courses do exist, there are about four of them worth naming, and together they run under six hours. That's the whole formal curriculum for image and video prompting as far as I can find it. Everything past that six hours is vendor documentation plus your own render count, and I'd argue the render count is where the actual skill lives.

This piece is the syllabus side. If what you want is a shopping comparison of the general text courses with prices and what each one skips, that's a separate piece at prompt engineer course and I'm not going to duplicate it here.

The Four That Actually Cover Generation

I went through the DeepLearning.AI catalogue and pulled the stated lengths off their own listings. These are the ones that touch images or video rather than chat.

Three of those four total five hours and thirty-four minutes. Add the fourth and you're at fourteen hours and forty-one minutes for the whole set, which is less than a single Udemy bootcamp and covers considerably more ground on the visual side.

The one I'd actually point people at is the agents course, and not for the reason it's marketed. The interesting idea in it is automatic evaluation of generated output, meaning the machine judges its own renders against a standard and reruns. That's the exact habit people skip when they're prompting by hand, and seeing it built as software makes the point better than being told.

Worth noting what none of them are. There's no course I found that teaches the vocabulary layer, meaning which words move which part of a picture, because that's model-dependent and it dates within months. That part you have to build yourself, and the way you build it is described in ai image prompt.

The Subscription Math, Briefly

Learn Prompting, which is a free-plus-paid platform rather than a marketplace, prices at $0 for the free tier, $21 a month billed annually for Plus, and $25 per seat monthly for teams of two to ninety-nine. Their free tier includes access to free courses and certificates of completion on all of them, which is unusual, since the certificate is normally the thing behind the paywall. Thirteen courses are listed on the plan page.

Twenty-one dollars a month annually works out to $252 for the year. If the only thing you wanted from any platform was the five and a half hours of generation-specific material listed above, you'd be paying something like forty-five dollars an hour of instruction, which is a strange way to think about it and I'm doing it anyway because nobody else does.

Set that against what the same money buys in renders. On a hosted API at the low quality tier, $252 is thousands of images. On a machine you own it's electricity. I run generation locally on an M4 Pro, so my own per-attempt cost rounds to nothing, and that's coloured everything I believe about this, which I'd rather say than pretend to neutrality.

The Skill Tree, Because Syllabi Aren't Ordered By Dependency

Courses are ordered by what's easy to teach. Skills have prerequisites. Those aren't the same order and it's why people bounce.

Level zero is understanding that the model is captioning rather than obeying, so you describe the picture that exists rather than instructing a person. Everything else fails until this lands.

Level one is the slot habit, meaning subject, lighting, shot, setting, style, ratio, exclusions, always in that rough order and always all present. At this level you're not choosing good values yet, you're just not leaving fields blank.

Level two is vocabulary, which is knowing what fifteen or twenty specific terms actually do because you rendered them yourself. This is the slowest level and the one no course covers. It is also, in my experience, the one where people either stall permanently or become genuinely good.

Level three is controlled comparison. One variable at a time, two renders per variant so you're not crediting a word for a seed. This is the skill that makes level two accumulate rather than reset.

Level four is consistency, meaning a fixed style block plus a swapped subject, seed discipline, and reference-based methods where identity matters. You cannot do this before level two because you don't yet know what belongs in the fixed block.

Level five, if it exists, is motion. Video prompting sits on top of all of it and adds an axis the models are visibly weaker at.

The mistake I see most often is people at level zero buying a level four tutorial, because the level four output looks impressive on a thumbnail. It doesn't transfer. There's no shortcut past the vocabulary layer and it's boring, which is exactly why it isn't sold.

Ten Questions, And If You Can Answer Them Skip The Course

Not a quiz with a score. A diagnostic. Answer these honestly and you'll know whether you need instruction or just hours.

Can you name the difference between distance and focal length as prompt controls? Do you know what happens to composition when you change aspect ratio, as opposed to what happens to the file? Can you name three lighting phrases and predict which one produces the moodiest result? Do you know why "flawless skin" makes a portrait less believable? Can you say what a seed does and when to lock it? Do you know which of your prompt's clauses is weighted most heavily? Can you explain why running a changed prompt once tells you nothing? Do you have a fixed style block written down outside a chat window? Can you name the failure that happens when a prompt has too many descriptors? And do you know what your own keep rate is, roughly, on images you'd publish?

If you answered eight or more, no course on any of these platforms is going to teach you much, and your bottleneck is render volume rather than instruction. If you answered three or fewer, the structure genuinely helps and I'd start with the shortest thing you'll finish rather than the best thing you'll abandon.

The last question is the one people find uncomfortable. Mine sits somewhere around one in four across everything, and worse on faces. Not knowing your own number means you have no baseline, so you can't tell whether a change helped.

Reading A Course Page In Two Minutes

Since new courses appear constantly and anything I list will age, here's the filter I run instead. Five checks, all doable from the sales page without buying.

Does the curriculum name specific models? A syllabus that says "AI image generators" rather than naming the thing it teaches against is either very old or written by somebody who hasn't run it recently. Generation courses date faster than text ones because the parameter surface keeps moving.

Does it show outputs? A course about pictures with no pictures in the curriculum listing is a bad sign, and I've seen several.

Is there an evaluation module? Anything at all about how you'd know the result got better. If the entire syllabus is techniques with no measurement, you're buying a vocabulary list.

Is the stated length believable? Seven hours of video is seven hours of video. Twenty-two hours usually means a lot of watching someone type. Longer is not better here and I'd weight a ninety-minute course you'll finish above a twenty-hour one you won't.

Is there a date anywhere? Last updated, a model version, a screenshot with a recognisable interface. Undated material in this field is a risk you're taking on.

Two of the four courses in my table pass all five comfortably. I'd rather hand you the filter than the ranking, because the ranking is wrong within a year and the filter isn't. Video-specific instruction is the thinnest slice of all of this, and the practical version of it is in ai video prompts.

What No Course Teaches, And Probably Can't

Three things, and they're the three that matter.

Knowing whether the output got better. Anthropic's prompting documentation opens by assuming you already have success criteria and a way to test against them, which quietly says the technique list was never the hard part. No course hands you a real standard, because doing that means giving you a real task with real stakes attached.

The second is when to stop. Every course teaches adding to a prompt. None teach subtraction, and past a certain density descriptors stop competing and start averaging, which is how you end up with a soft picture and no single word to blame.

Third is your own defaults. After a few hundred renders you develop habits, some good and some superstitious, and the only way to find the superstitious ones is to delete a clause you're sure about and render anyway. I've deleted phrases I'd defended for months and seen no change at all.

What Learners Ask

Is there a free ai prompt engineering course that covers images? The DeepLearning.AI short courses have been free to take during their platform beta, and the vision models one is ninety-two minutes. I'd start there rather than paying for anything.

Do I need to know Python? For the agents course, yes, meaningfully. For the rest, no. If you're prompting by hand in a web interface, code isn't the missing piece.

Are certificates worth anything here? For the image and video side I'd say no, since nobody is hiring on the strength of one and the portfolio is the evidence. For general AI literacy inside a company that reimburses training, different calculation entirely.

How long until I'm good? Depends entirely on render count rather than calendar time. I'd guess a few hundred deliberate one-variable renders is where things click, and that's an evening a week for a couple of months, not a weekend.

What about the free vendor documentation? Better than most courses, honestly. Google publishes filled-in templates for photorealistic scenes, stylized illustration, text in images, product shots, and more, and OpenAI documents their model's stated limitations. Both are free and neither is a course, which may be why nobody recommends them.

Should I learn text prompting first? Not necessarily. They're less related than the word prompt suggests, and the deeper text technique layer is in chatgpt prompt engineering if you want it.

The Ratio I Would Aim For

The vision models course, then close the browser and render two hundred images.

That ratio is the point. Ninety minutes of structure against several evenings of attempts, rather than the other way round, because the structure is small and the vocabulary is large and only one of them can be handed to you. If you want a wider map of how prompting differs across text, image, and video, it's in chatgpt prompts.