AI Video Upscaler: Turn Low-Res Clips Into Sharp 1080p or 4K

SquishyFile's free video upscaler lets you increase video resolution by 2x or 4x, so a 480p, 720p or 1080p clip comes out as sharper HD, 1440p or 4K. Most tools upscale video online on a remote server. I built this one to run inside your browser instead, so you don't upload the file anywhere, and the upscaled video has no watermark.

Small videos go through an AI model that rebuilds fine detail. Larger ones go through FSR, a fast upscaling shader. The tool picks the right engine for you, then shows a before/after slider so you can check the difference before you download.

How to upscale a video

What resolution will my video be?

The output is always your source resolution multiplied by the scale you pick. Here's what the common sizes turn into (16:9 video):

Upscale video to 4K: use 2x on a 1080p source (it lands exactly on 3840×2160), or 4x on a 540p source.

Upscale 720p to 1080p: choose 2x. You'll get 1440p, which is sharper than 1080p and plays perfectly on any Full HD screen or platform.

Upscale 480p to 1080p: 2x gives 960p, just under Full HD. For a true 1080p-or-higher result, use 4x.

Vertical phone videos follow the same math on their longer side, so a 720×1280 portrait clip at 2x becomes 1440×2560. Bigger isn't automatically better, though. Pick the smallest scale that reaches the resolution you need; 4x on an already large video creates a huge file and a long wait with little visible gain.

How the AI video upscaler works

When you add a video, the tool reads its height and routes it to one of two engines. Anything 720p or smaller goes to an AI model. Anything taller goes to FSR. I split it at 720p because an AI model has to run on every single pixel of every frame, and above that size it gets too slow to be practical in a browser, while FSR still does a good job on footage that already has plenty of detail.

The AI model I use is IMDN (Information Multi-Distillation Network), in its real-time variant. It's a tiny network, around 100,000 parameters, trained on the DIV2K set of high-resolution photos to predict the fine detail a low-resolution frame is missing. Most super-resolution models run every layer at full size. IMDN first shrinks the frame to half resolution, runs its six distillation blocks there, and only expands to the final size at the very end. That trick is what makes it fast enough to run on every frame of a video inside a browser tab. It works on the full RGB image, so edges and colors get rebuilt together rather than patched back in afterwards.

The model is trained for 2x. For 4x, the tool runs it twice in a row. Between the two passes it clamps the first result back into the normal color range, because the model slightly overshoots on hard edges, and a second pass would amplify that overshoot into visible halos. The frames stay on your graphics card the whole time through WebGPU: the video frame goes in, both passes run, and the finished frame goes straight to the encoder without being copied through the CPU. If your browser doesn't support WebGPU, the tool falls back to your CPU through WebAssembly: same result, just much slower.

FSR (FidelityFX Super Resolution 1.0) comes from AMD and was built for games. It enlarges each frame in two passes: the first follows edges so lines stay clean instead of turning into stair steps, and the second sharpens the result to bring back contrast. It needs no training data and runs very fast on the GPU, which makes it a good fit for 1080p and larger video. If you want the full story, I wrote a separate guide on what FSR is and how it works.

After the frames are upscaled, the tool encodes them into an H.264 MP4 at a high quality setting and adds the audio back as AAC, so the file plays on phones, TVs, editors and social platforms. It also fixes the orientation of phone videos in the pixels themselves, so portrait clips stay upright in every player. When the job finishes, the result shows the old and new resolution and which engine ran.

Can it enhance video quality?

Yes, when your video looks bad because it's too small. Most video quality enhancer apps bundle several fixes behind one button. I built this tool to do one of them well: resolution. It adds pixels, rebuilds the fine detail a low-res frame is missing and smooths out the blocky, pixelated look, so a small clip holds up on a TV or a big monitor.

Here's what it improves and what it leaves alone:

• Improves: resolution (2x or 4x, to 1080p, 4K or higher), the sharpness of edges and small text, and the pixelated look of low-res footage.
• Doesn't change: noise and grain, shaky footage, color and brightness, or frame rate. Motion blur and missed focus stay too, because the camera never recorded that detail.

So if your clip is 480p or 720p and looks soft on a bigger screen, this is the right fix. If it's already 1080p but dark, noisy or shaky, you need a different kind of editor, and upscaling it will only give you a bigger file with the same problems.

Videos that upscale best

These are the clips where upscaling can improve video quality the most:

• Old videos recorded in 480p or 720p on older phones and cameras
• Screen recordings and tutorials, where upscaling makes small text and UI edges easier to read
• Anime and cartoons, whose clean lines and flat colors upscale very cleanly, especially with anime mode on
• Low-resolution downloads that look pixelated on a TV or large monitor
• Clips you need to re-upload at 1080p or 4K for a platform or a client

Blur from a missed focus or fast motion is a different problem, and upscaling only makes it bigger. This guide to blurry video helps you tell which kind you have, and I cover whether video upscaling actually works in more depth on the blog.

How to upscale anime and cartoons

Anime needs different treatment from real-life footage. A camera shot is full of soft gradients, skin texture and noise, while a cel-animated frame is mostly flat color held together by thin, dark line art. Viewers judge anime quality almost entirely by those lines: when they're crisp and black, a 480p episode can look clean on a big screen; when they're fuzzy and grey, even a 1080p rip looks washed out.

That's a problem for general AI upscalers, mine included. IMDN learned from photos, so it rebuilds anime edges the way it would rebuild a photo: slightly soft, with lines that come out a little grey instead of solid black. Ordinary sharpening doesn't fix it either. It adds bright and dark halos around every stroke, which on flat anime colors looks far worse than on a busy photo.

So I added an Anime / cartoon checkbox under the scale picker. Tick it, and after the upscale the tool runs an extra pass on your graphics card that treats each frame like line art. It does three things, all on brightness only, so it never adds color fringes:

• Deblurs edges without halos. It sharpens, then limits every pixel to the brightest and darkest values already around it, so edges get steeper but never glow or ring.
• Darkens real lines only. It picks out dark strokes thinner than a small window and deepens them. A plain edge between two colored areas, like a white shirt against a wall, gets no fake outline.
• Thins lines back down. Upscaling makes every stroke fatter along with the frame. This step pulls pixels away from line centers, so lines look drawn at the new resolution instead of stretched.

Every step scales with the upscale factor, because a line that was 2 pixels wide in the source is 4 pixels wide after 2x and 8 after 4x. Anime mode works with both engines: on 720p and smaller videos it runs right after the AI model, and on larger videos it replaces FSR's own sharpening step, since it does a cleaner job on line art. It needs WebGPU; on the CPU fallback the checkbox has no effect.

A few tips for anime and cartoons:

• Leave it off for live action. On real footage the line-darkening turns hair, wrinkles and shadows into drawn-looking strokes.
• Use it for any flat-color art, not just Japanese anime: Western cartoons, 2D game footage, animated explainers and motion graphics all benefit.
• Old DVD anime at 480p is the classic case. Use 2x for a quick lift, or 4x if you want it at 1080p or above on a TV.
• A 1080p episode to 4K takes 2x on the FSR path plus anime mode, which is fast and keeps lines sharp at 3840×2160.
• Check the slider on a scene with faces and hair. That's where line quality shows first.

Tips for the best results

• Use the cleanest copy you have. The original file upscales better than a version that's been re-uploaded or sent through a messaging app, because every re-compression throws detail away.
• Start with 2x. Try 4x only if 2x doesn't reach the size you need.
• Test a short clip first. A few seconds is enough to judge the result on the before/after slider before you commit to a long video.
• Use Chrome. In my tests, Chrome ran clearly faster than Firefox on the same file.

How long does upscaling take?

The tool processes every frame on its own, so time grows with video length, source resolution and scale. A 30-second clip at 30fps is 900 frames; at 4x, each of those frames becomes sixteen times bigger before it gets encoded. A short clip at 2x usually finishes quickly, while a long video at 4x can take a while.

I tested it on a MacBook Air M3 with 16GB of RAM. Short clips ran smoothly at both 2x and 4x. Longer videos made the laptop warm, which is normal for a fanless machine under sustained GPU load. A desktop with a dedicated graphics card will be faster; an older laptop may be more comfortable with shorter clips and 2x.

Upscaling vs compressing a video

They do opposite jobs. Upscaling raises resolution and makes the file bigger. Compressing makes the file smaller so it fits an upload or email limit. If your upscaled video ends up too large to share, run it through the video compressor, or compress it to an exact size for Discord, email or WhatsApp. Upscale first, then compress, never the other way round: compressing first throws away detail the upscaler could have used.

Frequently asked questions