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.

v0.7.0

What's new

• Balanced and sensitive detection now recheck detector-backed remove ranges frame by frame, so brief subtitle gaps remain visible instead of being carried forward from coarse samples.
• WipeEngine.plan() now reports structured prepare, detect, refine, persist, and complete progress through the same callback model used by run().
• Plan generation now supports cooperative cancellation, rejects concurrent work on the same engine, and propagates callback failures without leaving the engine busy.
• Fast mode and fallback-only tracks keep their coarse temporal behavior and record that limitation in plan warnings.

Verification

• 205 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 was recorded at 355e7d6 and passed all gates; the frame 361 subtitle gap and aggregate false-removal ratio both measured 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.
• Temporal WipePlans require the built-in frame-based backend; file-based external backends accept only full-video static plans.

v0.6.0

What's new

• Added WipePlan, a source-bound JSON and NPZ contract for reviewing cleanup decisions before loading an inpainting model.
• Added temporal remove and keep tracks so masks apply only during the frame ranges where each target is present.
• Added safety defaults that keep persistent top overlays unless the request explicitly selects them for removal.
• Added --plan to the CLI and plan support to the structured Python SDK while preserving existing remove_text() and process() calls.
• Updated the local Web UI to show track actions and time ranges, toggle whole tracks, and execute the reviewed plan with precise masks.
• Hardened plan validation for source identity, mask assets, track segments, path traversal, and symlink escapes.

Verification

• 183 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 passed with the accepted WipePlan metrics and documented temporal-resolution limits.
• Wheel and source archives passed content checks and a clean-environment 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.
• Temporal WipePlans require the built-in frame-based backend; file-based external backends accept only full-video static plans.

v0.5.0

What's new

• Added a structured Python SDK with request, result, progress, cancellation, and stable error contracts.
• Made WipeEngine reusable across sequential jobs so one loaded model can serve a long-running worker or batch.
• Added a public Inpainter registry and runnable examples for integrating third-party video-repair backends.
• Switched the base package to headless OpenCV and kept ONNX, PyTorch, OCR, Web, and ProPainter support optional.
• Added a packaged ProPainter adapter without redistributing its separately licensed source code or model weights.
• Hardened wheel and source archive contents with isolated install checks and explicit package verification.

Verification

• 101 tests passed on Windows.
• Scoped ruff checks passed.
• Wheel and source archives passed content checks and a clean-environment install smoke test.

Notes

• Python 3.10 or newer is now 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.
• ProPainter remains an optional external integration with its own license and setup requirements.

v0.4.0

What's new

• Added a local-first Web UI via �ideowipe serve.
• Added the �ideowipe[web] extra for FastAPI/Uvicorn-based local usage.
• Added a review-first browser flow: upload, preview detected targets, confirm cleanup, and download.
• Preserved original audio in downloaded MP4 outputs.
• Added Reset handling that releases stale preview jobs instead of leaving the local server busy.
• Updated English and Chinese READMEs with Web UI screenshots.

Verification

• 70 tests passed on Windows.
• Scoped ruff checks passed.
• Wheel package check confirmed the Web server and HTML UI files are included.

Notes

• This release is local-first and single-user. It does not add hosted SaaS, authentication, multi-tenant queues, or cloud LLM calls.
• PyPI publishing is intentionally out of scope for this release.

v0.3.0

What's new since v0.2.0

Features

• Pluggable Inpainter protocol — src/videowipe/inpainters/ introduces a whole-video Inpainter protocol (InpaintJob/InpaintOutcome) and a name-based registry. STTN's three-stage backend contract (encode/transform/decode) is now a private implementation detail, not a public surface. Select models with --model (default sttn).
• ExternalInpainter + shell-injection fix — run_external now uses shlex.split + an argv list with shell=False instead of f-string + shell=True, eliminating shell injection via filenames. external is registered as a named inpainter.
• ProPainter as named registry entry — --model propainter --propainter-dir selects ProPainter. scripts/propainter_wipe.py resolves its source dir from --propainter-dir > VIDEOWIPE_PROPINTER_DIR > default (was hardcoded), switched to argparse, and writes a fixed inpaint_out.mp4 output (dropping the glob fallback).

Fixes

• Shell injection in external adapter — filenames with shell metacharacters were previously interpreted by the shell; now passed as literal argv entries.
• Version alignment — __version__ bumped to match pyproject.toml.
• Eval test PYTHONPATH — eval_clean_detection subprocess tests no longer fail with ModuleNotFoundError in environments without an editable install.

registry.names() now returns ['external', 'propainter', 'sttn'].

Full Changelog: v0.2.0...v0.3.0

v0.2.0

What's new since v0.1.0

Features

• videowipe clean intent-driven workflow — auto-detect removable regions from a video with --preview, --confirm, --target, --region, and --intent flags
• Frequency-map detection pipeline — samples frames, builds per-pixel text frequency map, extracts connected components, classifies by position and content
• Subtitle band fallback — catches bottom/top subtitle regions missed by the main pass
• --detect-mode presets — fast / balanced / sensitive with tuned sample count, consistency threshold, and fallback behavior
• --ocr text recognition — optional RapidOCR integration for reading detected text crops, helps classify timestamps and watermarks
• External model adapter — --external-command runs any CLI-based inpainting tool through the benchmark harness
• Docker delivery — CPU and GPU images, GHCR publishing via GitHub Actions CI
• Quality/performance benchmark — benchmark.json output, scripts/benchmark_pipeline.py, IoU metrics in eval script

Fixes

• Adaptive DBNet input size — widescreen (16:9) frames no longer compressed to 640×640 square, preserving bottom subtitle detection
• Smart preview frame selection — preview now uses the frame with the most detected text instead of always picking the first sampled frame
• FFmpeg pipe replaces cv2.VideoWriter — libx264 CRF 18 output instead of mp4v
• Manual DBNet fallback — high-level OpenCV API returning empty results falls back to manual post-processing
• Docker build fixes — quoted package spec, explicit --target builds, PR build verification

README

• Auto-detection accuracy showcase with multilingual preview images
• ProPainter comparison section with before/after images
• Bilingual GitHub About description
• Removed overstated model comparison conclusions

Full Changelog: v0.1.0...v0.2.0

v0.1.0

First release of videowipe. Model weight for subtitle removal.