Almost every guide to AI video generation is a guide to picking a model. Which one has the best physics, the longest clip, the cheapest credit. That comparison is real, and it is also the least important decision you will make, because the models converge every few months and the gap between them closes faster than you can act on it.
The decision that actually determines whether your video is watchable is made before any model runs: what you ask it to do at all.
The people getting reach from AI video are not the ones generating the most. They are the ones generating the least.
Here are the five rules we use, why each one is true, and what each one is arguing against.
Rule 1: Generate the stills. Move the camera. Don't generate the video.
This argues against: "describe your video and AI makes it."
The most expensive thing you can do with a video model is ask it to invent motion that a camera move would have produced for free.
A slow push into a good photograph reads as cinematic. So does a lateral drift, a pull back that opens a space out, a focus pull that moves attention from foreground to background. None of these require the model to hallucinate a single new pixel, because the image is not changing — the frame is. That is what a camera operator does, and it has carried film for a century.
Composed camera motion has three properties that generation does not. It is deterministic — the same input produces the same output, every time, with no reroll lottery. It is fast, because it is a filter graph and not a diffusion process. And it is cheap, by roughly an order of magnitude.
Reserve actual generation for the shots where the subject has to move: a person breathing, steam rising, fabric settling, water pouring. Everything else is a camera problem wearing a generation problem's clothes.
If you can describe the shot as a camera move, it is a camera move. Only the pixels that must change should cost money.
This is also why image-to-video overtook text-to-video as the serious creator's default. Text-to-video is the gateway; image-to-video is the upgrade. Starting from a still you control means the composition, the colour, the product, the face and the branding are all decided before the model is involved — and the model's job shrinks to something it can actually do reliably.
Rule 2: Buy motion at the hook, and almost nowhere else.
This argues against: "animate every scene."
Attention in short-form video is not distributed evenly, and it is not close. The opening seconds decide whether the rest of the video is seen at all — completion rate is the dominant distribution signal on every major platform, and completion is mostly determined before the third second.
So the marginal value of a generated shot is steeply front-loaded. The marginal cost is flat: a model charges the same whether the clip opens the video or sits at second eighteen where a third of your audience has already gone.
One generated shot at the top. Stills with camera motion everywhere after. If you want a second generated shot, you need a reason better than having the budget for it.
There is a second benefit, and it may be the larger one: the less generated footage is on screen, the less chance a viewer catches that any of it was generated. Restraint is both the cheaper choice and the safer one.
Rule 3: Lock identity, not endpoints. This is the whole morph.
This argues against: "morphs are trending, do a morph."
The morph is genuinely the format of the moment. Outfit changes, before-and-afters, room transformations, product reveals — two images and a model computing the motion between them. The common workflow is now well established: generate the subject, generate a matched variant, feed both in as first and last frame, let the model interpolate.
The morph is also, simultaneously, the most recognisable visual signature of AI slop. Both things are true at once, and the variable that decides which one you made is identity preservation.
An outfit change on a person who is unmistakably still the same person is a trend. A person dissolving into a handbag, growing an extra arm, or arriving at the end frame as somebody subtly different is the thing that gets an ad dragged within the hour. The audience cannot always articulate what went wrong, but they clock it instantly.
How to actually prompt one
Two techniques carry most of the difference.
Generate a composition-locked end frame. The end image should match the start on everything except the one thing that changes: same subject, same framing, same camera height, same lighting, same background. Most morph failures are not motion failures at all — they are end-frame failures, where the model was handed two images that disagree about more than one variable and had to invent a bridge across all of them at once.
Prompt the mechanism, not the destination. Describe the physical process of the change rather than the state you want to arrive at. "The fabric falls and resettles" produces motion. "Turns into a ballgown" produces a cross-fade, a smoke effect, or a ghostly double exposure where both states are briefly visible at once — the model's way of admitting it doesn't know how to get from A to B, so it dissolves instead.
If your before and after would fail a "same person?" test on a freeze frame, you have made slop. Fix the end frame, not the motion prompt.
For the specific case of a service-business transformation reel — the highest-converting use of this format — the structure matters as much as the morph. We covered it separately in how to make a before-and-after reel from photos.
Rule 4: Chain short shots. Never ask for one long take.
This argues against: "look how long clips can be now."
Clip-length limits are the headline spec of every model release, and they are close to meaningless as a quality signal. Single-pass generation tops out around fifteen to twenty seconds across the current generation of models, and quality begins degrading well before the ceiling. Past roughly thirty seconds, character appearance drifts, lighting shifts, and the camera stops obeying its own logic.
This is why experienced creators chain several short, internally consistent shots rather than forcing one long generation — and why anything genuinely long is stitched, not generated.
The real development in 2026 is that shot planning moved into generation. Several models now take a multi-shot brief and schedule the cuts themselves — describe a four-shot sequence and get four shots back with continuity between them, rather than four separate generations you have to reconcile. That is a meaningful improvement, and it does not change the underlying advice: consistency is far easier to defend across a cut than across a long take.
Decide where your cuts go before you generate anything. A cut hides a seam. A long take advertises one.
One hard-won detail
If you are mixing generated shots into a sequence of stills taken from the same source image, keep the generated shot standalone. Nesting an animated clip inside a run of frames from the same photograph is the most reliable way we have found to produce a truncated, mistimed sequence — the timing model has to reconcile two different authorities on how long the shot should be, and one of them loses. That is a finding from our own render pipeline rather than a published one, but it has cost us enough times to be worth stating plainly.
Rule 5: The payoff is variants, not volume.
This argues against: the entire "10× your content output" pitch.
This is the rule the data has moved on, and it is the one worth internalising before any of the others.
The pitch for AI video has always been volume — post ten times as much for the same effort. That worked for roughly eighteen months, and then the audience arrived at the other side of it. Consumer enthusiasm for AI-generated creator content fell from 60% in 2023 to 26% in 2025 as feeds filled with what viewers began calling slop. iHeartMedia's own research found 90% of its listeners — including the ones who actively use AI tools themselves — want their media made by humans. Platforms have started demoting the obvious cases.
But the speed that enabled the flood is the same speed that enables something much more valuable, and almost nobody uses it that way.
Generate five versions of your opening shot. Watch all five. Ship one. That is a workflow no human editor could have afforded five years ago, and it produces work that is indistinguishable from someone who tried very hard — which is precisely the bar the backlash has set. The audience is not objecting to AI. They are objecting to visible carelessness, which AI happens to make very cheap.
If AI let you post five times as much this month, you used it wrong. It should have let you post the same amount, five times better.
And watch the file before you post it
Do not trust a tool's description of what it made. Open the video. Wrong duration, missing audio, captions that never rendered and music that starts four seconds late are all common, all silent, and all invisible until the post is live and somebody else sees it first.
What the rules cost
The practical shape of a reel built this way, using our own render economics as the worked example:
| Shot | Approach | Relative cost |
|---|---|---|
| Opening / hook | Generated motion — the subject moves | ~10× |
| Body shots | Stills with composed camera motion | 1× |
| Reveal | Still, held, slow pull | 1× |
| Close / CTA | Still, detail shot | 1× |
One generated shot in a five-shot reel roughly triples the cost of the render and buys you the only three seconds that decide whether the other four shots are seen. Two generated shots roughly double it again and buy you very little. That ratio, not the model choice, is the economics of AI video.
We should disclose the obvious: Poppify is built to exactly this shape, so we are not neutral about it. The reasoning above is the reasoning that produced the product, rather than the other way round, and the cost breakdown in the real cost of AI video generation shows the same maths against the general-purpose tools.
Frequently asked
Should I use text-to-video or image-to-video?
Image-to-video, for anything where the output has to look like a specific thing. Text-to-video invents the subject, which means it invents your product, your premises and your face. Starting from a still you control removes that whole category of failure.
Is AI morphing still worth doing, or is it played out?
Worth doing, with the identity rule applied. The format has not fatigued — transformation content still has unusually high save rates — but careless morphing is now instantly legible as AI, which it was not a year ago. The craft bar moved; the format did not die.
How long should an AI-generated video be?
Short. Most vertical short-form performs best somewhere between roughly ten and thirty seconds depending on platform, and generated footage degrades past about twenty in a single pass anyway. The constraints happen to agree. Platform-by-platform numbers are in our video length breakdown.
Can I make a whole video with one AI tool?
Rarely, and the people producing the best work don't try. Multi-tool workflows are now standard — one tool for the stills, another for the generated shot, a third for assembly, captions and audio. The exception is composition tools that handle assembly natively, which is the category most small businesses actually need; we broke the three categories down in the buyer's guide.
Will AI video hurt my reach?
Obvious AI video will. Platforms have begun demoting it and audiences have begun reporting it. Video where AI did the assembly on footage and photographs you actually own does not read as AI to anyone, because it isn't — the intelligence went into the edit, not into inventing the subject. That distinction is the whole game now.
How many AI videos should I post per week?
The same number you would post without AI. Cadence is set by what your audience will tolerate, not by what your tooling can produce — and posting more than about five times a week measurably hurts engagement for most small accounts. The cadence data is here.
Related reading
- Pillar: AI Video Generator for Small Business — the three categories of tool, and which one composes reels from photos you already have.
- The Real Cost of AI Video Generation in 2026 — why per-clip pricing understates a finished reel by 5–10×.
- How to Make a Before-and-After Reel from Photos — the four-beat structure for the highest-converting transformation format.
- TikTok Video Length: the Data-Backed Sweet Spot
- How Many Social Media Posts Per Week for a Small Business