Pick your aspect ratio before you generate a single clip, not after you’ve already got footage you like. That one decision, made in the wrong order, is why so much AI-generated video looks soft, cropped, or oddly composed the moment it lands on a phone screen. Cropping a 1920x1080 (16:9) frame down to a 608x1080 vertical slice for 9:16 throws away roughly 68% of the original horizontal pixel data, according to Hailuo AI’s 2026 aspect-ratio guide. You’re left with about a third of the image you thought you had.I found this out the annoying way, mid-project on Lost Garden. I’d generated a batch of wide establishing shots for a trailer cut, all in 16:9, because that’s the ratio I default to for anything cinematic. Then a teammate asked for a vertical version for socials. I cropped the center out of the same clips instead of regenerating, because regenerating felt wasteful. The vertical cut looked muddy. Faces lost detail. Fine texture in the background, the stuff that sells a shot as “real,” went soft first. It wasn’t a rendering problem. It was a math problem I’d created three steps earlier by not deciding the final format before I hit generate.What Aspect Ratio Should You Use for AI-Generated Video?The honest answer is: whatever ratio your final platform actually plays back in, decided before you write a single prompt. Not after you see how a shot turns out.16:9 is still the cinematic default: YouTube, festival submissions, a desktop-first trailer, anything meant to feel like a film rather than a phone video.9:16 is the mobile-native format for TikTok, Instagram Reels, and YouTube Shorts. It’s not a smaller version of 16:9, it’s a different composition, built around verticality instead of horizontal negative space.1:1 still shows up in feed placements and some experimental or art-house work, though it’s the least common of the three in 2026.None of these is “correct” in the abstract. The correct one is whichever your finished piece will actually be watched in. Decide that first, and the rest of the workflow gets simpler, not harder.Why Does Cropping After Generation Look So Bad?Because the math is brutal and most people never do it before they hit publish. A standard 1080p frame is 1920x1080 pixels. To fit that into a 9:16 vertical container without stretching, you keep a centered 608x1080 slice and throw away the rest. That’s about 32% of the original frame surviving the crop, and roughly 68% gone.The damage shows up in specific, predictable places:Faces first. Skin texture and micro-expression are high-frequency detail, and high-frequency detail is exactly what a 68% crop throws away.Backgrounds turn to mush. Anything meant to read as “real,” gravel, fabric weave, distant signage, loses its edges and starts to look painted.Motion gets choppier. Cropped footage that’s then upscaled back to a delivery resolution has less real information per frame, so movement looks less smooth even at the same frame rate.There’s a second, sneakier version of the same mistake: feeding a square (1:1) reference image into an image-to-video tool and asking it for a 16:9 output. The model doesn’t have real pixel data for the extra width, so it invents some. Hailuo AI’s guide calls this “hallucinated edges,” which is a polite way of saying the AI is guessing what belongs at the sides of your frame. You usually can’t tell just from the thumbnail. You can tell the moment someone views it full-screen.Aspect ratio is a directorial decision, not a delivery setting you fix at the end. Treating it as an afterthought is what turns an expensive-looking clip into a soft, cropped one.Do the Big AI Video Models Even Handle Aspect Ratio the Same Way?No, and the gap is widening, which is exactly why building your own ratio-first habit matters more than betting on any single tool’s current feature set.Google’s Veo 3.1 update added native 9:16 output at up to 4K resolution (3840x2160), meaning it can generate a true vertical clip instead of cropping a horizontal one down, according to Google’s own release documentation and independent coverage from outlets like CineD. That’s a real fix for one specific failure mode. It doesn’t fix the same problem on every other tool in your stack.Meanwhile, the tool matrix under your feet keeps shifting. OpenAI’s Sora web and app experiences went dark on April 26, 2026, with the developer API following on September 24, 2026, confirmed on OpenAI’s own help center. If your workflow depended on Sora’s specific aspect-ratio handling, that dependency now has an expiration date. A workflow built around a decision you make yourself, before generation, survives a tool’s shutdown notice. A workflow built around one tool’s quirks doesn’t.This is a pattern worth internalizing: every model handles ratio, resolution, and reference-image matching a little differently, and those specifics change every few months. The fix isn’t memorizing each tool’s current behavior. It’s deciding your target ratio before you open any of them.How Do You Build an Aspect-Ratio-First Workflow?A short checklist, run once per project before you generate anything:Name the final platform first. Not “I’ll figure it out later.” YouTube trailer, TikTok cutdown, festival submission: each one has a native ratio, and that ratio is your answer.Match your reference image to your target ratio. If you’re doing image-to-video and your seed image is square but your target is 16:9, you’re inviting hallucinated edges before you’ve generated a single frame of motion.Never switch ratios mid-sequence. A batch of clips meant to cut together should share one ratio from the first generation to the last. Mixing ratios inside a sequence breaks visual continuity in a way that’s hard to fix afterward.If you genuinely need two platforms, generate twice. It costs more up front and saves you the muddy crop later. Treat it as two separate deliverables with two separate prompts, not one clip stretched to cover both jobs.This is also why, when I rebuilt the planning stage for ScreenWeaver, I made the shot plan ask for a target format before a single prompt gets written, not after a shot comes back and someone has to decide what to do with it. It’s a small, boring gate. It’s also the single workflow change that stopped the “wait, why does this look soft on my phone” conversation from happening on every project.The Lost Garden trailer eventually got regenerated properly in 9:16 rather than cropped. It took longer than the crop did. It also didn’t look like a screenshot of a bigger video.FAQWhat’s the best aspect ratio for AI-generated video in 2026?There isn’t a single best ratio. 16:9 fits cinematic and desktop-first delivery, 9:16 fits TikTok, Reels, and Shorts, and 1:1 still covers some feed placements. The right one is whichever platform the finished clip will actually play on.Can I just crop a 16:9 AI video down to 9:16 after generating it?You can, but you’ll lose roughly 68% of the original horizontal pixel data doing it, based on the standard 1080p crop math. It’s a usable emergency fix, not a workflow.Does using a square reference image cause problems for 16:9 video generation?Yes. Image-to-video tools have to invent pixel data for the extra width when the reference image doesn’t match the target ratio, which is where soft or hallucinated edges come from.Is resolution more important than aspect ratio for AI video quality?They’re different problems. Resolution is about pixel density, aspect ratio is about composition and how much of that density survives a crop. Getting the ratio right before you generate protects the resolution you already paid for.If aspect ratio is the decision that’s been costing you regenerations, it’s worth locking it in before your next shoot, not after you see the first cut.