Free AI upscaler comparison: UpgoLAB vs Upscayl vs Video2X vs Topaz

Q: Can I upscale video without a GPU?

Yes, by using a cloud-based tool. UpgoLAB processes on cloud RTX 5090/3090 GPUs, so any device with a browser works, including phones and low-end laptops. The tradeoff is a 30 second limit per video file.

Q: What resolution should I upscale to?

At most 2x per axis. Past roughly 2x the model stops reconstructing and starts inventing detail, which reads as waxy skin and fake text. A 540p source at 2x lands at 1080p, which is enough for anything watched on a screen.

Four tools that do the same job in four different ways. Which one fits depends on whether you have a GPU, how long your clip is, and whether you are willing to install anything.

Quick answer

No GPU, no install, short clip → UpgoLAB (browser, free during launch)
Have a GPU, photos only → Upscayl (open source, local)
Have a GPU, long video → Video2X (open source, local)
Professional long-form work, willing to pay → Topaz Video AI

Comparison table

	UpgoLAB	Upscayl	Video2X	Topaz Video AI
Runs in browser	Yes	No, desktop	No, desktop	No, desktop
Needs local GPU	No	Yes	Yes, Vulkan	Yes, strong GPU
Free tier	Yes, unlimited during launch	Yes, open source	Yes, open source	No
Photo support	Yes, up to 4 MP	Yes	No	Yes
Video support	Yes, up to 30 s	No	Yes, unlimited length	Yes, unlimited length
Content-specific presets	8 modes	Several models	No	Some
Watermark on free	No	No	No	n/a
Works on phone	Yes	No	No	No
Setup time	None	Install + model download	Install + build	Install + licence

What each one is actually good at

UpgoLAB — when you have no hardware

Runs on cloud RTX 5090/3090 GPUs, so the device you are holding does not matter. That is the whole point: a phone photo of a grandparent, or a 20-second clip pulled from a messaging app, can be fixed without buying a graphics card.

The constraint is real and worth stating plainly: 30 seconds per video file, 4 MP per photo. For longer files you have to split first (commands below).

Eight presets exist because content degrades differently. Compression blocks need a different treatment than sensor grain, and anime line art needs the opposite of what night footage needs.

Upscayl — best local photo upscaler

Open source, wraps Real-ESRGAN and similar models. Excellent for photos if you already own a capable GPU. No video support at all. A common choice for people restoring scanned prints.

Video2X — best local video upscaler

Free and open source, handles video of any length and also does frame interpolation. Requires a Vulkan-capable GPU and some patience with the build. The documentation is good; the setup is not trivial for non-technical users.

Topaz Video AI — the paid benchmark

Best-in-class for long-form professional work, with model choice per clip and mature handling of motion. The reason to avoid it is cost, and the reason to consider it is a 40-minute timeline you need to deliver today.

The part nobody tells you: order matters more than tool

Most guides say enlarge first. That makes the model learn the compression blocks and enlarge them too. The sequence that holds up:

Denoise, targeting block artifacts rather than sensor grain
Enlarge at most 2x per axis
Inspect at 100% on the hair edge, not at fit-to-window
Grade last

Past roughly 2x the model switches from reconstructing to inventing. Faces go waxy and small text turns into strokes that look like letters but are not.

Compression noise is not grain. Compression imposes a regular block grid; sensor grain is random. A generic denoiser treats the grid as noise and leaves it in place, then the upscaler amplifies it. Fixing the order changed results more than switching models ever did.

Splitting a clip over 30 seconds

Cut at keyframes so the source is not re-encoded:

ffmpeg -ss 00:01:12 -to 00:01:38 -i input.mp4 -c copy segment.mp4

Batch split:

ffmpeg -i input.mp4 -c copy -f segment -segment_time 30 -reset_timestamps 1 seg-%03d.mp4

Rejoin without a second encode:

printf "file '%s'\n" seg-*.mp4 > list.txt
ffmpeg -f concat -safe 0 -i list.txt -c copy output.mp4

Results at 100% crop

Each of these is the model output, not a marketing render:

UpgoLAB portrait mode result, 100 percent crop showing hair edge detail

UpgoLAB natural mode result on a compression-damaged source

UpgoLAB texture preservation mode, fabric and skin grain retained

UpgoLAB denoise then 2x enlargement, block artifacts removed

All eight content modes

Each preset targets a different kind of damage. These are the same photos the tool ships as demo material:

01 Natural
02 Ultra Crisp
03 Social Denoise
04 Portrait Studio
05 Anime 2D
06 Cinema Vivid
07 Clear Reality
08 True Texture

Frequently asked

What is the best free AI upscaler that runs in a browser?

UpgoLAB runs entirely in the browser with no install and no local GPU. Photos up to 4 MP, video clips up to 30 seconds, enlarging to 2K or 4K. Free during launch, no watermark.

Is there a free alternative to Topaz Video AI?

For short clips, UpgoLAB is free and needs no GPU. For long video on your own hardware, Video2X is free and open source but needs a Vulkan GPU. Neither matches Topaz on long-form work, but both cover the common case.

Can I upscale video without a GPU?

Yes, with a cloud tool. UpgoLAB processes on cloud GPUs, so any browser works including phones. The tradeoff is a 30 second cap per file.

What resolution should I upscale to?

At most 2x per axis. A 540p source at 2x lands at 1080p — enough for anything watched on a screen. Pushing 4K from a 480p source produces a larger file with the same real information.

Why does my upscaled photo look worse than the original?

Usually one of three things: denoising after enlarging (blocks already amplified), a single model used for both night grain and line art (opposite requirements), or judging quality at fit-to-window instead of 100% on an edge.