Aspect Ratio, Duration, Resolution, and FPS: The Constraints That Bite

By now you can pick a tool and read the leaderboard, you can write a structured prompt, and you know to approve a still frame before you animate it. So you're getting good clips. Today is about the four boring numbers that quietly wreck good clips, the ones nobody warns you about until you've already burned the credits. Aspect ratio, duration, resolution, and frame rate. None of them are glamorous. All of them will bite you on a deadline. The good news is they only bite when you decide them late, and the whole point of this episode is to decide them first.

Let me start with the one that hurts the most when you get it wrong, because you usually can't fix it after the fact.

Aspect ratio is a decision you make before you write a word

Aspect ratio is just the shape of the frame, the width compared to the height. You already know these shapes even if you've never named them. Landscape, wider than it is tall, is sixteen by nine, the shape of a television or a YouTube video on a laptop. Vertical, taller than it is wide, is nine by sixteen, the shape of your phone held upright, which is what TikTok, Instagram Reels, and YouTube Shorts all want. Square is one to one, an even box, which shows up in some social feeds. There's also the older four by three, more boxy, and the very wide cinemascope shape, twenty-one by nine, that you see on movies, though most generators don't offer that one yet.

Here's the thing that trips people up. Aspect ratio is a generation setting, not an editing setting. You tell the model the shape before it makes anything, and it builds the whole frame to fit that shape. It is not something you sensibly fix afterward by cropping. The current hosted tools all expose a ratio picker, and the menus differ. Google's Veo generates landscape and vertical natively. ByteDance's Seedance offers a long list, square, landscape, vertical, the boxy ratios, and the wide ones. Kling gives you square, landscape, and vertical. MiniMax's Hailuo covers the common ones plus a two-to-one. The exact menu doesn't matter for today. What matters is that you choose from it deliberately, matched to where the video will actually play, before you spend a credit.

So why is choosing late so expensive? Picture the classic mistake. You generate a beautiful landscape hero shot, sixteen by nine, your subject centered with the room around them, supporting action off to the sides, real cinematic width. It's gorgeous. You lock it. Then the client says, oh, we also need it vertical for TikTok. So you crop the landscape frame down to a vertical slice. Watch what happens. A landscape frame cut to vertical throws away roughly forty percent of the picture or more. The wide context you carefully built vanishes. The supporting action on the sides is just gone, cut clean out of frame. Your subject, who looked great centered in a wide shot, might now be half out of the slice, an ear missing, hands cropped off. And because you're keeping only a narrow vertical strip of the original pixels, the resolution of that strip is a fraction of what you started with, so it looks softer too. You lose the composition, you lose the context, and you lose sharpness, all at once. There's a good walkthrough of exactly this trap in a guide to adapting videos across aspect ratios, and the upshot is simple, cropping is not converting.

A model that's generating natively in vertical knows it has a tall, narrow canvas. It places your subject and your motion to fit that canvas from the first frame. You get a real vertical shot, not a salvage job. So the rule is, generate native to the shape you need. If you need both landscape and vertical, you've really got two shots, and the cleanest answer is usually to generate each one natively rather than crop one out of the other.

When you genuinely have to serve both from one generation, there's a framing trick worth knowing. Frame so the vertical cut works first. If your subject and the important action all fit inside the central vertical window, then pulling back out to the full landscape frame will still hold together, and every shape in between those two will work as well. The same source describes this, compose for the tightest crop, and the looser crops come along for free. It's a compromise, not a substitute for native generation, but it'll save you on a multi-platform job.

One more practical layer on vertical, the safe zone. When your video plays on TikTok or Reels, the platform stacks its own interface on top of your frame. Across the top sits the username and the audio tag, a couple hundred pixels of it. Across the bottom sit the caption, the buttons, the call to action, three hundred-odd pixels. And the right edge has the like and share column eating into your picture. So even though you fill the whole nine-by-sixteen frame, the safe area, the part that won't get covered, is a narrower column down the middle, weighted toward the upper third, because that's where a viewer's eyes land when they're holding a phone. Keep your subject and any on-screen text in that central band. A vertical video safe-zone guide lays out the exact margins if you want them, but the habit is what counts, don't put anything you care about where a button will land on it.

And know your destinations, roughly. YouTube on a desktop is landscape, sixteen by nine. YouTube Shorts, TikTok, and Instagram Reels are all vertical, nine by sixteen, and conveniently the same vertical file, ten-eighty by nineteen-twenty, plays natively on all three. As of early 2026, Meta even unified Facebook and Instagram Reels into one vertical safe zone, so a single vertical cut covers both. Decide the destination, and the ratio decides itself.

The square and the boxy ratios have their places too, so don't ignore them. Square, the one-to-one box, still shows up in some feed placements and in ad slots where the platform wants a tile that reads the same whether someone's scrolling on a phone or a desktop. The older four-by-three, slightly boxy, turns up when you're matching archival footage or a retro look on purpose. And the very wide cinemascope shape, when a tool offers it, buys you a deliberately filmic, letterboxed feel for a trailer or a title sequence. The point isn't that you'll use these often. It's that aspect ratio is an expressive choice as much as a technical one, and the safest move is still the same, pick the shape on purpose, up front, and generate to it, rather than discovering halfway through the edit that the shape you have isn't the shape you need.

Duration: there's a wall, and it's lower than you think

The second number is duration, how long a single clip can be. New people assume they can ask for a minute. You usually can't. Most hosted generators hand you something in the four-to-ten-second range per single clip, with a few stretching toward fifteen. The exact caps move constantly and vary by model and tier, so verify before you promise a client anything, but the shape of reality is short. Some tools advertise longer numbers through an extend feature, where the model tacks another few seconds onto the end of an existing clip, and a couple of the frontier models claim longer single generations, but treat the big numbers with suspicion and test them on your own shot, because advertised maximums and good-looking maximums are not the same thing.

Why the wall? It isn't an accident, and it isn't really a limit to fight. It's there because of how these models drift. Generated video loses coherence the longer it runs. The industry term is temporal drift, which just means consecutive frames stop agreeing with each other. A face slowly changes shape. A hand gains or loses a finger. A pattern on a shirt wanders. A background object slides somewhere it shouldn't. There's a clear explanation of this in a piece on solving temporal drift in AI video, and the gist is that small errors early on compound, each frame drifting a little further from the last, until the illusion falls apart. As a rough feel, quality often starts to slide somewhere past the five-to-eight-second mark. The model's memory of where it started decays over time, and detail collapse is what you see.

So the duration cap is the model protecting you from its own drift. Which means the right response is not to hunt for whoever offers the longest single clip. The right response is to stop thinking in single long renders and start thinking in short clips you stitch together. You generate a clean six-second shot, then you take the last frame of that clip and use it as the start frame of the next one, exactly the start-frame move from last episode, and you keep going. Because clip two literally begins on the frame clip one ended on, they flow together. That's keyframe chaining, and it's how professionals build something thirty or sixty seconds long out of a model that only makes eight seconds at a time. We're giving chaining its own full episode next, because there's craft to doing it without drift creeping in across the seams, so I'll just plant the flag here. The duration wall is real, and chaining is how you walk around it instead of through it.

The thing not to do is max out the duration slider on a single generation and hope. That's how you get a clip that's lovely for four seconds and falls apart for the last four, and you've paid for all eight.

A word on the extend features, because they're easy to misread. Several tools let you push past the base cap by extending, Kling's version adds a few seconds at a time, for example, and each extension costs more credits. Extending is genuinely useful, but it's the same drift problem wearing a friendlier face, because you're asking the model to keep going from where it was, and the further it goes the more it wanders. So treat an extend button as a short reach, not a license to build a minute out of one continuous generation. For anything substantial, deliberate chaining, where you choose the hand-off frame yourself and write a fresh motion prompt for the next shot, gives you more control than letting the model autopilot another ten seconds. The destination of the chain is something you steer. The destination of a long auto-extend is something you discover.

There's also a planning consequence here that's easy to miss until it bites. If your finished piece is going to be a chain of short clips, then continuity becomes your job, not the model's. The lighting, the color, the subject's look all have to carry across the seams, and that's a craft we'll build up over the next several episodes. For today, the takeaway is narrower, when you plan duration, you're really planning your cut points. Decide where the shots break before you generate, so each clip is a deliberate beat in the sequence rather than an arbitrary eight-second chunk you'll have to rescue in the edit.

Resolution: draft cheap, finish sharp

The third number is resolution, how many pixels make up the picture, which translates directly to how sharp and detailed it looks. Let me say these for the ear. Seven-twenty p, sometimes called HD, is the web-safe, fast, cheap tier. Ten-eighty p, full HD, is the standard for YouTube and social, and that ten-eighty by nineteen-twenty vertical file is the native size for TikTok, Reels, and Shorts. Two K is a step above that, less common in these tools. And four K, sometimes called Ultra HD, is the cinema-grade tier, roughly four times the pixels of ten-eighty p, and it's the slow, expensive one.

Here's the workflow that saves you the most money in this entire episode, and it's dead simple. Draft at low resolution, finish at high resolution. When you're still figuring out the shot, whether the composition works, whether the motion is right, whether the timing lands, generate at seven-twenty p. It's faster and it's cheaper, and you're going to throw most of these away anyway, so don't pay for pixels you'll discard. Only once the shot is locked, once you've picked the take you're keeping, do you spend the credits on a high-resolution final pass. One finishing write-up frames it as a clear rule, run a stack of fast cheap drafts to lock composition and timing, then a single high-quality render for delivery, and notes that this can cut your iteration cost by more than half compared to rendering everything at full quality. There's no point upscaling or high-rendering a take you might still revise. Lock the creative first, then buy the quality.

And if your final delivery needs to be sharper than the tier you generated at, there's a finishing path called upscaling, which uses a separate tool to intelligently add resolution to a clip after it's made. The best-known one is Topaz Video AI, a paid desktop app that runs a whole library of specialized models to upscale, clean up noise, sharpen, and recover detail, and it can turn a clean ten-eighty p clip into a sharp four K delivery. It runs on the order of a few hundred dollars a year. Most upscalers give you a faithful mode that enlarges while preserving your original look, and a more aggressive mode that invents extra texture, and for client work you usually want the faithful one. We'll spend real time on upscaling and finishing in a later episode, so for now just know the path exists, you don't have to generate at four K to deliver at four K.

Why does higher resolution cost more in the first place? Because the model is doing more work, holding a bigger image in memory, running slower, sitting in a longer queue, and the platforms charge more per second for the higher tiers on top of that. A clip's price scales with resolution and with duration together, so a long four K clip is the most expensive thing you can ask for, and a short seven-twenty p draft is the cheapest. Knowing that shape is what lets you spend deliberately.

Let me put rough numbers on it so the scaling is concrete, with the loud caveat that prices change monthly and vary by tier, so verify before you quote any of this to a client. Most of these tools bill in credits or in cents per second of video. As reference points people have published, a turbo tier on one major tool runs around five credits per second of output. A ten-second clip at the full ten-eighty p professional tier on another can run into the couple-hundred-credits range. One value-focused model advertises generation down around a nickel per second at its lower resolution. The exact figures aren't the lesson. The lesson is the shape of the bill. Per second, times a resolution multiplier, times however many rolls you do. That middle term, the resolution multiplier, is the one you control completely by drafting low, and that last term, the number of rolls, is the one the draft-cheap habit shrinks. You can't change what a model charges. You can absolutely change how many expensive seconds you buy.

So picture the math on a real job. A vertical social clip, ten-eighty by nineteen-twenty, eight seconds, thirty frames. You might run ten fast seven-twenty p drafts to nail the composition and the timing, each one cheap, then a single full-resolution final render of the winner. The ten drafts together can cost less than one full-quality render would, and you end with one delivered clip. Now picture the chained version, a thirty-second sequence built from three ten-second shots. You draft each of the three cheaply until the chain holds together, then pay for three final renders, one per shot. The expensive part is three final passes, not thirty seconds of trial and error at full quality. That's the cost-per-finished-clip mindset doing its job. Same finished length, a fraction of the spend, because the drafts lived in the cheap tier and only the keepers got the expensive treatment.

Frame rate: the look of motion

The fourth number is frame rate, frames per second, which is how many still images flash by each second to create the motion. This one's less about sharpness and more about feel. Twenty-four frames a second is the cinema standard, it has a slightly soft, filmic, dreamlike quality, and it's what gives movies their movie-ness. Thirty frames a second is the broadcast and online standard, cleaner and more neutral, the look your brain reads as television or normal web video. Sixty frames a second is smooth and crisp, the look of live sports and video games, and it's also what you want if you're going to slow footage down into slow motion, because you've got extra frames to stretch.

Higher frame rates cost more, for the same reason higher resolution does, more frames means more to generate. So you don't always need to generate at sixty. Like frame rate can be changed after the fact, with a step called frame interpolation, where a tool invents brand-new in-between frames to raise the rate, turning twenty-four or thirty up to sixty or higher, or stretching a clip into clean slow motion. Topaz does this with dedicated interpolation models, and there are standalone tools for it too that can take a clip up to sixty or even a hundred and twenty frames a second. So if a model outputs at twenty-four and you need thirty for broadcast, or sixty for a smooth social slow-mo, you can interpolate up afterward rather than paying to generate high frame rates natively. For most social delivery, thirty frames a second is the safe default. For a filmic look, twenty-four. Reach for sixty when motion smoothness or slow motion is the point.

There's one frame-rate trap that catches people building chained sequences, and it's worth flagging now. If you generate one shot at twenty-four frames and the next at thirty, then cut them together, the motion feel jumps at the seam, and it reads as a glitch even to a viewer who couldn't tell you why. So pick one frame rate for a whole piece and hold it across every clip in the chain, the same way you hold one aspect ratio and one look. Frame rate is a project-level decision, not a per-clip one. Decide it once, up front, alongside the other three numbers, and apply it everywhere. The interpolation tools are there to rescue a mismatch after the fact, but it's cleaner and cheaper to just not create the mismatch.

How the four bite together, and the pre-flight that stops them

Individually these are just settings. Together they're a planning sequence, and running that sequence before you generate is the whole skill. Here's the order.

Start from the destination. Where does this play? That single answer drives the next two. The destination sets your aspect ratio, vertical for TikTok and Reels and Shorts, landscape for YouTube and desktop, and it sets your length expectations. From there, make a duration plan. What's the longest single shot you actually need, and if it's longer than the model's cap, where will the cuts fall, which is to say, plan the chain. Then pick a draft resolution, which should almost always be the lowest tier, seven-twenty p, for fast cheap iteration. Then decide your final resolution and frame rate, which for most social work is ten-eighty p at thirty frames, and four K only for hero or feature pieces. And finally, if your final is higher than your draft, plan the upscaling step in, so it's a decision, not a panic at the end.

The money lesson threaded through all of that is don't draft expensive. Don't generate four K drafts. Don't max the duration on every roll. Run a batch of cheap, low-resolution, short drafts to lock the composition and the motion, and only then spend on one high-quality, correctly-shaped, final render. That same finishing guide pegs the draft-then-final habit at roughly a sixty percent saving over rendering everything at full quality. This is the cost-per-finished-clip mindset from episode one, made concrete. You're not trying to make any single generation cheap. You're trying to reach a finished, delivered clip for the fewest total credits, and most of your rolls should be cheap drafts that never ship.

Let me give you the one pre-flight workflow to copy. Step one, turn the brief into a ratio. TikTok ad, that's vertical, non-negotiable, decide it now. Hero shot for both YouTube and Instagram, that's really two shots, plan the primary natively and the secondary natively too if you can. Step two, write the prompt with the constraints baked in. Say the shape and the length right in the prompt, vertical nine by sixteen, eight-second clip, so the model frames for it from the start, which ties straight back to the prompt anatomy from episode two. Step three, draft batch. Set resolution to the lowest tier, generate three to five variants, and judge them for the things that go wrong, drift, the subject staying intact, and whether the framing actually fits the destination. Step four, creative lock. Pick the take that works and don't move on until composition and motion are right. Step five, the final render, regenerate the approved take at delivery resolution in the correct aspect ratio. Step six, post-process only if needed, upscale to four K if the client needs it, interpolate the frame rate up if you need smoothness or slow motion. And step seven, export to platform spec. For vertical delivery across TikTok, Reels, and Shorts, one export setting covers all of them, ten-eighty by nineteen-twenty, nine by sixteen, an MP4 file using the H two six four codec, at thirty frames a second.

The pitfall, named so you'll recognize it

Let me make the main pitfall vivid, because you will face exactly this. An agency generates a gorgeous landscape hero shot for a campaign. Talent centered, landscape context around them, supporting actors in frame, motion across the wide canvas. Twenty seconds, ten-eighty p, locked and approved, everyone's happy. Then the client adds, almost as an afterthought, we also need it vertical for TikTok. The team, on a deadline, crops the landscape file to vertical instead of regenerating. The landscape context disappears. The supporting talent is cut clean out of the frame, and the story loses its beats. The resolution of the cropped strip collapses, because you're keeping only a narrow column of the original pixels. The client rejects the crop, correctly, and now the team regenerates the shot from scratch at vertical, having paid for it twice and blown the timeline.

The fix is entirely upstream, and it's a single question asked before you generate anything. Does this shot need to work vertical? If yes, generate it native at nine by sixteen from the start, or frame the landscape conservatively enough that a vertical center-crop still holds. Write the answer into the brief in plain words, primary delivery vertical for Reels and TikTok, secondary landscape for YouTube. Cost the shot once, by deciding its shape before the first credit, not after the client sees it.

Where this sits, and what's next

So the four numbers, decided in order, before you generate. Destination sets the aspect ratio, the ratio you generate native rather than crop into. Duration has a low wall, so you plan short clips and chain them instead of chasing one long render. Resolution is draft cheap and finish sharp, with upscaling as the bridge to four K. And frame rate is the feel of motion, twenty-four for film, thirty for normal, sixty for smooth, changeable after the fact with interpolation. Run that pre-flight and you stop discovering constraints the expensive way.

This sets up the next two episodes directly. The duration wall is exactly why the next episode is keyframe chaining, taking the last frame of one clip into the next to build sequences longer than any single generation allows. And the draft-cheap, finish-sharp habit is what the later upscaling and finishing episode picks up, where we get into Topaz, frame interpolation, and grading in real depth. One more tie-back worth making, the leaderboard skill from episode one applies here too, when you bench the top models on your own shot, bench them at your aspect ratio, your duration, and your resolution, because a model that's great at landscape eight-second clips might not be the one you want for vertical fifteen-second ones. Test the constraints you'll actually ship under.

And if you take just one habit from this whole episode, make it this. Before you spend a single credit, say the four numbers out loud, the shape, the length, the draft resolution, and the frame rate. Four seconds of planning that saves you a re-render every time. The creators who deliver on deadline aren't the ones with the best prompts. They're the ones who decided these four boring numbers before they hit generate.