ClapClip AIClapClip AI
Download

Measurements

Benchmarks

Numbers from this machine, on these files, measured this way. Every study below publishes the hardware, the test images and the scoring method next to the result, because a timing without its conditions is not a measurement.

What every study includes

  • The machine. CPU, RAM, GPU and operating system, plus the parts we could not pin down — for example when the GPU adapter in use was not logged.
  • The files. Pixel dimensions, format and how the input was produced, including any degradation we applied ourselves.
  • The method. What the timer starts and stops on, which metric and parameters were used, how many runs, and which statistic is reported.
  • The limits. What the result cannot support: sample size, synthetic inputs, and metrics that score fidelity rather than what looks good.

Published studies

2× Image Upscaling: AI vs. Lanczos on Two Test Images

Two reference photos were downscaled by half and compressed to JPEG at quality 50, then upscaled 2× with the Windows app and with OpenCV Lanczos. Both outputs were scored against the untouched originals.

Result: PSNR 27.331 vs 27.240 dB and 27.088 vs 26.506 dB; SSIM 0.8546 vs 0.8301 and 0.7622 vs 0.7333.

Does a Face Swap Stay Inside the Face? A Pixel Check

Each face-swap output was compared against a no-op control from the same SDK, then every changed pixel was classified as inside or outside a hand-marked face region.

Result: 100% of changed pixels fell inside the region; mean difference outside it was 0.0007 / 255 at 512 px and 0.0004 / 255 at 256 px.

Two studies so far. New measurements are added here as they are run, not on a schedule.

What we do not publish

  • Competitor comparisons. We have not run a controlled test against another product, so we do not rank ourselves against one.
  • Population accuracy.A handful of test images shows how the model behaves on those images, not a success rate for everyone’s photos.
  • GUI timings presented as inference time. Where a number includes model loading or process start-up, the method says so.

Re-run it yourself

Each study keeps its inputs, outputs and scoring script alongside the report, so a result can be reproduced instead of taken on trust. If a number here disagrees with what you see on your own machine, the conditions are the first thing to compare — different GPU, different source photo, different scale.

Try ClapClip for Windows