nadermx/backgroundremover

📋 Changes

• Fall back to `moviepy.editor` import when the moviepy-2.x top-level import is unavailable
• Use `Clip.resized()` (moviepy 2.x) or `Clip.resize()` (moviepy 1.x), whichever exists
• Shim `Image.ANTIALIAS` -> `LANCZOS` (removed in Pillow 10), which moviepy 1.x's resize references
• Relax dependency pin to `moviepy>=1.0.3`

📦 Highlights

• ProRes 4444 default: Transparent video output now defaults to `prores_ks` with `yuva444p10le` (10-bit color with alpha) instead of `qtrle`. Much better compression, color fidelity, and NLE compatibility (DaVinci Resolve, Premiere, Final Cut Pro). The old `qtrle` codec is still available via `--alpha-codec qtrle`.
• Smart worker scaling: Worker count (`-wn`) is now dynamically capped based on your actual GPU VRAM rather than a hard limit. The tool calculates per-worker memory (CUDA context + model + JIT trace + batch tensors) and prevents over-allocation that caused hangs.
• Mask threshold for sharp edges: New `-mt`/`--mask-threshold` option binarizes the alpha mask at a given threshold (0-255) for hard/crisp cutouts — ideal for cartoonish images, logos, and illustrations.
• Device diagnostics: Prints detected device, GPU name, and VRAM at startup so you can verify GPU acceleration is working.

✨ New CLI Flags

• | Flag | Description |
• |------|-------------|
• | `-mt 128` / `--mask-threshold 128` | Binarize mask for sharp edges (0-255) |

🐛 Fixes

• #181: Video processing hanging with high worker counts — workers now dynamically capped to what GPU VRAM can support.
• #122: Soft/blurry edges on cartoonish images — new `--mask-threshold` option for hard cutouts.
• #145: No indication of GPU usage — now prints device info (CUDA/MPS/CPU) with GPU name and VRAM at startup.

💥 Breaking Changes

• Default transparent video codec changed from `qtrle` to `prores_ks`. Output files will be smaller with better color but require ProRes-compatible playback. Use `--alpha-codec qtrle` for the old behavior.

📦 Upgrade

• ```bash
• pip install --upgrade backgroundremover
• ```

📋 Changes

• Video pipeline stability and compatibility improvements across transparent outputs.
• #44: Use exact detected frame rate to reduce alpha mask drift.
• #85: Explicitly map filtered video + optional audio so outputs aren’t “audio only.”
• #49, #86: Added --alpha-codec / --alpha-pix-fmt options to control alpha video encoding; lossless default remains qtrle.
• #98, #88, #152: More robust alpha merge pipeline and encoding options improve results on problematic videos.
• #168: Corrected mask usage in transparentvideooverimage so alpha is applied consistently.
• New CLI flags: --alpha-codec (auto, qtrle, prores_ks, libvpx-vp9) and --alpha-pix-fmt.
• README updated with alpha codec guidance and framerate override examples.

📋 Changes

• GUI application added (PR #175)
• Background image now optional, doesn't hang on stdin (PR #179)
• Better CLI validation with helpful error messages
• Improved multiprocessing error handling
• Warning for high worker counts

📋 Changes

• Fixed error running on Docker (#176)
• Resolved issue where Docker always downloads models (#105)
• Fixed video background remover not working in Docker (#120)
• Fixed transparent .mov files not being created (#135)
• Resolved video removal producing mask only (#134)
• Fixed invalid argument error when creating transparent mov files (#115)
• Fixed overlay transparent over image with "invalid value" error (#116)
• Fixed RuntimeError: operator torchvision::nms does not exist (#172)
• + 10 more