v0.8.0

What's new

• Long and high-resolution videos now use smaller inpainting segments by default, stream sampled detection frames, and avoid retaining full-resolution frames between model passes.
• Stable center watermarks can now include adjacent graphic marks, and VideoWipe masks them before STTN inference so the model does not reconstruct the removed overlay.
• DBNet detection reuses sampled results and skips redundant high-level inference after the manual path proves reliable, reducing balanced detection time without changing the selected tracks.
• Apple Silicon prefers the faster Torch backend when both Torch and ONNX Runtime are installed, and the benchmark script now records repeatable single-video timing, input identity, environment, and process RSS.

Verification

• 230 tests passed locally and in GitHub CI on Ubuntu with Python 3.10 and 3.13, plus Windows with Python 3.12.
• The formal detection baseline remained unchanged: remove Jaccard 0.239025, Boundary F 0.536385, keep coverage 0.0, and false-removal ratio 0.0.
• Wheel and source archives passed content checks, and the wheel passed the clean Ubuntu installed-runtime smoke test.

Notes

• Python 3.10 or newer is required.
• VideoWipe is distributed under GPL-3.0; review the license before embedding or redistributing it.
• PyPI publishing remains out of scope for this release; install from source or use the published container images.
• The default inpainting segment is now 25 frames; larger values provide more temporal context but increase compute and memory cost superlinearly.