Dataset Viewer
Auto-converted to Parquet Duplicate
scene
stringclasses
3 values
problem
stringclasses
2 values
input
stringclasses
2 values
output
stringclasses
2 values
method
stringclasses
1 value
before_image_url
stringclasses
3 values
after_image_url
stringclasses
3 values
page_url
stringclasses
3 values
result
stringclasses
3 values
neon shop signs
low resolution
360p
720p
general-x4v3 fp16 (AI enhance) vs normal resize
https://tohd.video/img/home-before.jpg
https://tohd.video/img/home-after.jpg
https://tohd.video/
Chinese characters come back with crisp, clean strokes
freckled face close-up
slight defocus
360p
720p
general-x4v3 fp16 (AI enhance) vs normal resize
https://tohd.video/img/unblur-before.jpg
https://tohd.video/img/unblur-after.jpg
https://tohd.video/unblur-video
Sharper eyes, eyelashes and nose ring; fine freckles can be smoothed out
children at a birthday party
low resolution
480p
1080p
general-x4v3 fp16 (AI enhance) vs normal resize
https://tohd.video/img/p…party-before.jpg
https://tohd.video/img/p…-party-after.jpg
https://tohd.video/upscale-video-to-1080p
Sharper eyes, hair and polka dots

WebGPU video upscaler tests

Measured results for webgpu-video-upscaler, an MIT-licensed video upscaler that runs Real-ESRGAN general-x4v3 with ONNX Runtime Web on WebGPU, inside a browser tab. It is the engine behind tohd.video. I'm on the team that builds it.

Files

  • comparisons.csv: three before/after frames from our own tests on free stock footage, scaled down or blurred on purpose and then upscaled. Each row has the problem, the input and output size, links to the "before" (normal resize) and "after" (AI enhance) frames, and the page where the sample is shown.
  • speed.csv: per-frame timings. All numbers come from one machine: an Apple M2 Max (38-core GPU) in a Chromium-based browser with WebGPU, on 8 Nov 2026. No other devices have been measured yet.

What the tests show

  • Low resolution (360p → 720p, 480p → 1080p): visibly sharper; text strokes, whiskers and eyelashes come back as clean lines.
  • Slight defocus: better. Motion blur: not fixed.
  • Heavy compression: slightly better; blocks are smoothed, lost detail does not come back.
  • fp16 gives the same output as fp32 (same PSNR in our test) and was 5 to 8 times faster at 360p.

Limits

One machine only, so speed on other GPUs, phones and integrated laptop GPUs is unknown. Output is capped at 1080p on the short side. The frames are hosted on tohd.video; the source footage is from Pexels.

Write-up with the full tables: https://tohd.video/open-source-video-upscaler

Downloads last month
51