← Back to the learning paths
Apply AI to your jobIntermediate → Advanced

Generative Media

Image, video and audio you can actually deliver — because a brief is not a lottery ticket.

There is a moment, usually in week three, when someone realises they can generate a beautiful image in twenty seconds and still cannot deliver the one the client asked for. That gap is the entire subject of this path. Generating is free and nearly instant; matching a brief — this product, this angle, this mood, this face again in the next six frames, at this size, by Thursday — is a craft, and it is not the same craft as writing a clever prompt. Prompt-only workflows plateau fast, and the plateau is visible: the results are gorgeous and generic, they all have the same plastic sheen, and none of them are what was ordered. The people who get past it stop treating the model as an oracle and start treating it as a camera with strange controls. They work with references, masks, structural conditioning, and above all with iteration loops that converge instead of wandering. This path teaches that. It also spends real time on the two things nobody wants to discuss at the demo: art direction, because the bottleneck is taste and taste is trainable but not downloadable; and rights, because the first serious client will ask where the image came from, whether that face belongs to anyone, and what you are prepared to sign. You will end with a portfolio and a measured number — usable assets per hundred generations — that tells you more about your skill than any gallery.

LevelIntermediate → Advanced
Estimated duration4-6 months
Phases3

What you will learn

How diffusion models actually generateStructural control: references, masks, conditioningConvergent iteration instead of random re-rollingArt direction, briefs and reference boardsCharacter, product and brand consistencyVideo motion and temporal consistencyVoice, sound design and audio generationRights, provenance, disclosure and client contracts

Prerequisites

  • A visual eye you are willing to have criticised
  • Basic image editing — layers, masks, colour
  • Comfort with prompting an AI model at a basic level
  • A real brief from a real person, even an unpaid one

Where it leads

  • Generative Media Artist
  • AI Art Director
  • Creative Technologist
  • Content production lead (agency or in-house)

Phases

Phase 1 — How the machine draws, and how to steer it

Stop rolling dice. Understand the mechanism well enough to aim it at a specific result.

Estimated duration · 4-5 weeks

Diffusion, seeds and why prompts plateau

A diffusion model starts from noise and removes it while your prompt whispers directions. That single fact explains almost every frustration: why the seed changes everything, why adding adjectives stops helping after a point, why the model gives you the average of a million similar pictures. Once you see it as guided denoising rather than as a machine that understands your sentence, you stop arguing with it and start steering it.

Topics covered

Noise, denoising steps and guidance strengthSeeds, determinism and reproducing a resultLatent space and why resolution is not just sizeWhy extra adjectives stop workingThe averageness problem: beautiful and genericCommon artefacts and what causes themModel families and what each one is good at

What you will build

  • Reproduce the exact same image twice from a locked seed, then change one parameter at a time and document what each one does
  • Take one gorgeous generic result and, without changing the prompt, make it specific using seed and guidance alone
  • Build a personal reference sheet of artefacts with the setting that caused each one

Control: references, masks and conditioning

This is where amateurs and professionals separate, and it has nothing to do with vocabulary. Professionals impose structure from outside the prompt: a pose, a depth map, a layout sketch, a masked region to regenerate while the rest stays untouched. A mediocre generation you can fix in three targeted passes beats a spectacular one you cannot repeat. Learn to fix rather than re-roll, because re-rolling is how a two-hour job becomes a two-day job.

Topics covered

Image-to-image and strength as a dial, not a switchInpainting, outpainting and surgical fixesStructural conditioning: pose, depth, edges, layoutReference images and style transfer honestly assessedUpscaling, detail passes and where they lieCompositing: the generated layer is rarely the final layerBuilding a convergent loop instead of a random walk

What you will build

  • Take a badly framed generation and fix the composition with masking only — no new prompt, no re-roll
  • Reproduce a given reference photograph’s exact composition with a completely different subject
  • Deliver one image that matches a written brief within three targeted iterations, and log every pass

Phase 2 — Direction, consistency and motion

Serve a brief, hold a look across a whole campaign, and make things move without them melting.

Estimated duration · 7-9 weeks

Art direction and killing the AI look

Everyone can spot it: the waxy skin, the impossible lighting, the symmetry nobody asked for, the composition that belongs to no one. That look is the model reverting to its average, and defeating it is an art direction problem, not a settings problem. You will learn to write a brief that constrains, to build reference boards that mean something, and to hold one visual point of view across twenty assets — which is what a client is actually buying.

Topics covered

Reading a brief and finding the constraint that mattersReference boards that direct rather than decorateLight, lens and imperfection as anti-generic toolsCharacter and product consistency across a setBrand systems: colour, framing, toneCultural specificity and avoiding the postcard clichéCritique: giving and surviving it

What you will build

  • Produce eight images that unmistakably belong to the same campaign, and have three people confirm it without being told
  • Take the same subject and deliver it in two visual directions a client could genuinely choose between
  • Show one before/after where you removed the AI look, and write down exactly what you changed

Video and audio: time changes everything

Video is not images in a row, and treating it that way is why so much generated footage looks like a melting dream. The moment you add time you inherit temporal consistency, motion that must obey physics the model never learned, and continuity across shots. Audio has the opposite trap: it is technically easy and ethically loaded, because a voice belongs to a person in a way a landscape does not.

Topics covered

Temporal consistency and why faces driftMotion, camera language and physics the model ignoresShot length, cuts and editing around weaknessesContinuity across shots in a sequenceVoice synthesis, cloning and consentMusic, sound design and the silence you forgetHybrid workflows: generated, filmed, and repaired by hand

What you will build

  • Cut a thirty-second sequence where the same character survives four shots without changing face
  • Rescue a failed generated clip in the edit — retime, crop, cut around the artefact — and show both versions
  • Produce a fully voiced spot, then write the consent and disclosure note you would attach for a client

Phase 3 — Rights, provenance and production

Deliver on a deadline, at a known cost, with an answer to every question a client’s lawyer will ask.

Estimated duration · 5-7 weeks

Rights, provenance and what you can sign

The first serious client will not ask whether your image is beautiful. They will ask who owns it, whether that face belongs to a real person, whether the style is traceable to a living artist, and whether you will indemnify them. Most creators discover these questions at signature, which is the worst possible moment. Provenance is a workflow habit, not a legal afterthought: if you cannot say where an asset came from, you cannot sell it to anyone who matters.

Topics covered

Who owns a generated asset, and why the answer variesLicences, commercial use and platform termsLikeness, voice and the consent nobody collectedStyle, imitation and living artistsProvenance records and asset traceabilityDisclosure: when to say it was generatedIndemnity clauses and what you should refuse

What you will build

  • Write a one-page provenance record for a delivered asset: sources, tools, references, human contribution
  • Draft the clause you would add to a client contract, and the one you would strike out
  • Audit your own portfolio and flag every asset you could not defend if challenged

The production pipeline

One image is a hobby. Forty variants across three formats, delivered Thursday, versioned so the client can compare round two with round one — that is the job. The number that measures you is usable assets per hundred generations, and it should rise every month. Track it, because a portfolio hides your failures while that ratio does not, and clients pay for reliability far more than for peak brilliance.

Topics covered

Usable assets per hundred generations, tracked over timeNaming, versioning and asset management that scalesBatch generation and selection disciplineCost per delivered asset, not per generationFormats, crops and deliverable specificationsClient rounds, feedback and scope creepKnowing when to photograph, illustrate or generate

What you will build

  • Deliver a full campaign kit from one brief: eight assets, three formats each, on a date you committed to in advance
  • Measure your usable-asset ratio on two projects a month apart and explain the difference
  • Produce one deliverable where generation was the wrong tool, and say so to the client with a better proposal

Questions

Will this replace photographers, illustrators and designers?

It replaces execution, not judgement — which is uncomfortable, because a lot of people were paid for execution. If your value was producing a competent image from a clear instruction, that value is collapsing and pretending otherwise helps nobody. If your value was deciding what the image should be, arguing with a client about why their idea is wrong, and holding a look across a campaign, you just acquired a very fast assistant with no taste. The winners in the next few years will be visual people who learned the tools, not tool people who hope taste is optional. It is not, and it takes years, and that is genuinely good news for anyone willing to put in those years.

Can I actually sell this work to clients, legally?

People do it every day, and the ones who sleep well are the ones who track provenance from the first pass. This path does not give legal advice and the rules differ by country and keep moving, so treat this as a working posture rather than a ruling. The practical stance: know which tool produced each asset and under what terms, never generate a recognisable real person’s face or voice without their written consent, be wary of styles traceable to a specific living artist, keep a record of your own human contribution because it is often what makes the work defensible, and be transparent with the client rather than letting them find out later. The creators who get into trouble are almost never the ones who asked the question early. They are the ones who assumed nobody would.

Do I need a powerful machine, and do I need to know how to draw?

A powerful machine helps and is not a gate — hosted tools will take you through most of this path, and renting compute for the heavy phases costs less than the time you would waste configuring a local setup badly. Drawing is a different matter. You do not need to render like an illustrator, but you do need the underlying literacy: composition, light, colour relationships, why one crop works and another does not. Without it you cannot judge your own output, and judging is the whole job now that producing is free. If you have never studied any of it, spend an hour a week on visual fundamentals in parallel with this path. It will do more for your results than any new model release.

Related paths

Want to walk this path with a coach and a cohort?

Discover the 212AY Academy →