laidline-bench

Robustness and calibration benchmarks for the Laidline watermarks: the code behind laidline bench, plus the calibration and replay tools that produced the shipped decoder thresholds. The laidline package itself does not depend on it; install it to re-measure a mark on your own media or to reproduce the published numbers.

pip install laidline-bench        # pulls in laidline[image,audio,video,text]
laidline bench image --corpus ./photos --n 50
laidline bench audio --corpus ./speech --modes streaming
laidline bench video --corpus ./clips --device cuda
laidline bench text  --corpus ./c4-shard.json

• Each suite embeds a watermark, applies a set of transforms (re-encoding, resizing, cropping, noise, codecs, text edits) and decodes again. The result is a JSON file with per-transform detection rates; --resume continues an interrupted run from its checkpoint, --out chooses where the result goes.
• --device cpu|cuda|auto picks the torch device (video also takes cuda:N). A GPU run uses strict float32 and records the numerics next to the result.
• Without laidline-bench, laidline bench exits with status 2 and prints the install command.
• The image suite's img2img regeneration sweep needs a Stable Diffusion pipeline: pip install laidline-bench[diffusers].
• Modules are importable as laidline_bench.image_robustness, laidline_bench.audio_robustness and so on; each has a main(argv) taking the same flags as the command.

The methodology and the numbers are in the robustness and benchmark pages of the documentation: https://laidline.com/docs/robustness.