Upscale an AI video clip or re-render at top resolution?

You have a clip you like, rendered small, and now it has to look good on a big screen. There are two ways to get there. Send the finished clip through a video upscaler, or go back to the video model and render the same shot again at the highest resolution it offers. They sound like two roads to the same place. They are not: one keeps your take and polishes it, the other trades your take for a sharper one that may move differently.

This post is the decision, question by question. The short version: if the motion is the thing you cannot lose, upscale. If the detail is the thing that is wrong, re-render.

What each road actually does

An upscaler takes your frames and makes them bigger, filling in detail the small render never had. The motion, the timing, the framing and every lucky accident stay exactly as they were. What it cannot do is fix anything that happened in the render: a warped hand stays warped, only sharper.

A re-render asks the video model to draw the shot again from your prompt, this time with more pixels to draw into. Detail comes from the model itself, so it is usually more coherent than detail an upscaler guesses. But it is a new take. Even with the same prompt, the same seed and the same first frame, a different resolution can change how things move.

Question 1: can you lose this take?

Be honest about why you like the clip. If it is the moment the dog turns its head at exactly the right beat, or the way the smoke curls once and settles, that is luck, and luck does not survive a re-render. Upscale it. If what you like is the idea, the composition and the light, and the motion is merely fine, a re-render costs you nothing you care about.

This is why the draft, lock, finish workflow finds the motion on cheap drafts and only then renders the final at full quality. When you plan ahead, the re-render is the finish. Upscaling is the rescue for the clip you did not plan to keep.

Question 2: how big are the faces?

Faces decide more of these calls than anything else. In a small render, a face that fills a third of the frame has enough pixels to be a real face, and an upscaler will sharpen it well. A face in a wide shot may be a handful of pixels: eyes and mouth are a smudge. An upscaler has to invent that face, and it may invent a slightly different one on each frame, so the person seems to change as they move.

Rule of thumb: close-ups upscale well, small faces in wide shots should be re-rendered. If the re-render keeps losing the face too, frame closer. The lens and focus guide has the words for a tighter shot.

Question 3: does it flicker?

Watch the small render at full screen before you decide, and look at fine texture: hair, grass, fabric, water, brick. If that texture already crawls or shimmers, an upscaler makes the shimmer bigger and sharper, and a creative upscaler can add its own on top, because invented detail is invented again on every frame. A re-render at a higher resolution gives the model room to draw that texture properly, and the crawl often goes away.

If the clip is calm and the texture is stable, upscaling is safe. Pick a faithful setting over a creative one for anything with skin, text-like patterns or a repeating surface.

Question 4: what does each cost?

Both roads are priced by the length of the clip, so a longer clip costs more either way. The difference is what else moves the price. A re-render at the top resolution is usually the most expensive way to make that clip at all, and you may need more than one try before the motion is right again. An upscale is a single pass over a clip you already own, and its price depends on the target size you ask for. Every model quotes the price for your settings on the Generate button before you run it, so compare the two quotes for your actual clip rather than guessing.

The cheapest path is often a mix: draft small until the motion is right, then re-render once at the top resolution. Reach for the upscaler when the take is irreplaceable or when the model you used has no bigger size to offer.

The video upscalers in the catalog

Five models in the catalog take a finished video and return a bigger one. They differ mostly in how much control they give you:

• [FLUX Video Upscale](/models/black-forest-labs/flux-3/video-upscale) has a switch between a source-faithful and a creative upscale, a scaling factor, and an optional prompt that guides the new detail. Leave the prompt empty for a neutral upscale.
• [Fibo Video Upscaler](/models/bria/fibo/video-upscaler) lets you pick the target resolution and the output format, including editing-friendly containers, which helps if the clip goes into a video editor next.
• [Crystal Video Upscaler](/models/clarity-ai/crystal-video-upscaler) asks for a target size in megapixels, handy when the delivery spec is a pixel count and not a named resolution.
• [Video Upscaler](/models) offers a short list of target resolutions and is the simple choice for a plain resolution bump.
• [Upscale V1](/models/runwayml/upscale-v1) has no settings at all: upload the clip, get it back bigger.

Each one caps how long an input clip may be, and the caps differ. Check the model's page before uploading a long clip, or cut it into parts.

If you re-render

Keep everything that made the take: the same prompt word for word, the same seed if the model has one, and the same first frame on an image-to-video model. Kling V4.0 image-to-video also takes a last frame, which pins where the shot ends. Then change only the resolution. If the new take drifts, add the motion you liked to the prompt in plain words, so it no longer depends on luck.

What neither road can do

An upscaler adds pixels, not frames: it will not smooth jerky motion, slow the clip down or make it longer. The timing words guide covers speed, which is set at render time. Neither road repairs one bad frame in a good clip either; there is no frame-by-frame retouch here. And no upscaler can recover a face or a logo the small render never drew, it can only guess one.